DISSERTATION DEFENSE

Author :  Vedant Khandelwal
Advisors: Dr. Amit Sheth
Date: Aug 5, 2026
Time: 08:00 am
Location: STB 529, AI Institute 
Link:   https://sc-edu.zoom.us/j/82296571999

Abstract
Neural and symbolic methods fail in complementary ways. A neural model can answer almost any question but cannot distinguish correct from plausible. A symbolic model says what correct means but covers only what its authors anticipated. Many ways to combine them exist, and no principle says which one a task needs; this dissertation supplies one.

The symbolic part fixes the space of allowed options, and the neural part ranks within it. What varies is where along the pipeline, from training data to returned output, that space gets fixed. A check decides whether a candidate satisfies a stated criterion without constructing it, whether a coloring is valid, whether a plan is executable, and whether an answer is correct. Placement follows from what a task's check settles, what it costs, and what it leaves open, not from the application. Retrieval admits no check at query time, since relevance is a human judgment, so the space is fixed at the input. Reasoning admits one that settles an answer outright, so it is fixed after generation. Planning admits an equally cheap check but has too few training domains, so it is fixed before learning, and again at generation, where an exact search keeps every step legal while the heuristic ranks.

Applied systems meet several signals at once and take no new rule, since each subproblem gets the placement its own check predicts. Removing each in turn shows what it contributes. The result is a decision procedure from how a task can be checked to where symbolic structure belongs, what it gains over a neural baseline, and when it stops paying.