Title: Towards reliability and controllability of machine learning models

Date: Monday, April 28, 2025
Time: 1:00pm - 3:00pm ET
Location: (hybrid) CODA C1308 Cabbagetown and Zoom https://gatech.zoom.us/j/98399801627?pwd=rN37t8hs0ErO38Us0xIOVywtzUbKaE…

Kartik Sharma
School of Computational Science and Engineering
College of Computing
Georgia Institute of Technology

Website: https://ksartik.github.io/

Committee:
Dr. Srijan Kumar, Advisor (School of Computational Science and Engineering, Georgia Institute of Technology)
Dr. Bo Dai (School of Computational Science and Engineering, Georgia Institute of Technology)
Dr. Chao Zhang (School of Computational Science and Engineering, Georgia Institute of Technology) 
Dr. Rakshit Trivedi (Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology)

Abstract:
Despite their remarkable capabilities, the widespread adoption of machine learning (ML) systems is hindered by their unpredictability and lack of user control. This dissertation addresses these challenges by advancing the reliability and controllability of ML models across three key directions:
1.    User-controllable generation: A projected sampling method, PRODIGY, is introduced for generating graphs/drugs from pre-trained graph diffusion models under hard structural constraints (e.g., molecular weight, valency) without retraining. To support domain-expert control in retrieval-augmented generation, a hypergraph-based retrieval method, OG-RAG is designed to ground query-relevant contexts in domain-specific ontologies.
2.    Measuring reliability: Practical reliability gaps are exposed in dynamic graph neural networks (GNNs) by formulating undetectable perturbations using a novel constraint and demonstrating novel vulnerabilities with an efficient PGD-based approach. A systematic framework, ELK, is also proposed to efficiently identify factuality gaps in large language models (LLMs) by estimating their knowledge using sentence embeddings.
3.    Robust and domain-specific architectures: By leveraging sociological theories of structural balance, SEMBA is designed to learn effective representations of nodes in dynamic signed social networks. Another contribution, MetSelect, enhances the current architectures of GNNs by learning to select a node-personalized representation layer to classify each node. 
Future work extends these directions to language and multimodal data, focusing on LLM robustness via adaptive system prompt mechanism, imitation learning agents with inner speech to better capture human behavioral diversity, and reinforcement learning-based token pruning for memory-constrained LLM reasoning.