Doctoral Speaking Skills Talk - Saranya Vijayakumar

— 12:30pm

Location:
In Person - Blelloch-Skees Conference Room, Gates Hillman 8115

Speaker:
SARANYA VIJAYAKUMAR , Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://svijayakumar2.github.io/index.html

Grounding Neural Inference with Satisfiability Modulo Theories: Leveraging Theory Solvers for Machine Learning

This presentation describes SMTLayer, a framework for incorporating Satisfiability Modulo Theories (SMT) solvers into deep neural networks. Unlike previous approaches that approximate solver behavior with differentiable relaxations, SMTLayer directly integrates non-differentiable SMT solvers into both forward and backward passes of network training. 

SMTLayer enables encoding rich domain knowledge as mathematical formulas within the network architecture. During the forward pass, the solver uses symbols from prior layers to construct inferences based on these formulas. In the backward pass, the solver informs network updates, driving it toward representations compatible with the solver's theory. 

Through experiments on visual arithmetic, algebraic equation solving, and natural language reasoning tasks, we demonstrate that models using SMTLayer require significantly less training data than conventional models, show greater robustness to covariate shifts, and learn naturally interpretable representations.  

Presented in Partial Fulfillment of the CSD Speaking Skills Requirement


Add event to Google
Add event to iCal