Year: 2021
Type: dissertation
Abstract: Program synthesis is the process of learning mappings between sets of inputs and outputs in a way that generalizes to new inputs. Contrary to deep learning in the gradient descent sense, the goal of program induction is not to 'converge' to a correct solution by performing gradient descent on millions of parameters - rather it is to generate and search for discrete programs that are expressed as combinations of a library of known 'concepts' that will solve the given problem. The goal of this thesis is to explore the portability of program induction onto the bioinformatics domain – specifically the problem of tumor grade prediction. Programs enumerated to predict tumor grade from a data set of colon cancer were up to 76% accurate when the library of primitives was limited to arithmetic, exponential and logarithmic operations. Further work will involve building in models for solving differential equations (another success was to induce Dreamcoder to discover the forward Euler method for solving PDEs), as well as building conceptual representations of n-dimensional spatial data such as images. data such as images.
Author: Khirwar, Madhav
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7025592793
