A Neural Network Solves, Grades & Generates University-Level Mathematics Problems by Program Synthesis
Despite the impressive achievements of deep neural networks (DNNs) in recent years, researchers generally believe these models are still not “intelligent” enough to solve advanced mathematical problems in calculus, differential equations, linear algebra, etc.
In the new paper A Neural Network Solves and Generates Mathematics Problems by Program Synthesis: Calculus, Differential Equations, Linear Algebra, and More, a research team from MIT, Columbia University, Harvard University and the University of Waterloo challenges this assumption, proposing a neural network that can solve university-level mathematics problems by turning questions into programming tasks, i.e. program synthesis.
The team says theirs is the first demonstration of a neural network capable of solving university-level mathematics problems, which it does by combining two recent innovations:
- Neural networks pretrained on text and fine-tuned on code, rather than pretrained on text alone.
- Novel techniques that automatically augment problems with context so neural…
