Research

Research Themes

• Conformal prediction and risk control
• Uncertainty quantification and decision rules in large language models (LLMs)
• Evidential/Bayesian deep learning for epistemic uncertainty
• OOD detection and robustness

Selected Ongoing Projects

• Evidential OOD detection using misclassification cost through regularization
• Adaptive risk control in LLMs
• Evidence-based model uncertainty quantification from input data quality

Selected Research Experience

Uncertainty Quantification in Language Models Using Adaptive Conformal Semantic Entropy

Designed an adaptive Semantic Entropy framework that estimates LLM uncertainty using dispersion of sentence embeddings over clustered generations, then inflates/adjusts uncertainty to improve conservativeness and mitigate hallucinations. The approach applies conformal calibration to obtain distribution-free, guaranteed prompt-level accept/abstain decisions and response-level prediction sets.

CEAR: Certified Ensemble Adversarial Robustness in Neural Networks

Proposing a hybrid ensemble defense combining variable Gaussian augmentation and temperature-scaled distillation with noisy logits and a robust weighted ensemble at inference. Extends randomized smoothing to ensembles to improve certified accuracy, robustness radius, and resistance to strong white-box attacks on image classification benchmarks.

SACP: Spatially-Adaptive Conformal Prediction in Uncertainty Quantification of Medical Image Segmentation

Developed a spatially-adaptive conformal prediction framework for medical image segmentation that augments standard CP with class-conditional calibration and distance-weighted nonconformity scores. Produces anatomically informed prediction sets that expand near critical tumor–vessel interfaces while preserving finite-sample coverage guarantees across multi-centre pancreatic tumor datasets.

Model Uncertainty Quantification in Deep Neural Networks Using Evidential Properties and Conformal Prediction

Designed conformal nonconformity scores that incorporate evidential and information-theoretic properties to quantify model (epistemic) uncertainty in deep classifiers. Generates adaptive prediction sets while maintaining empirical coverage of the ground truth.
Ongoing direction includes conformal risk control for deployment.

Certified Robustness in Deep Neural Networks Using Diverse Ensemble and Noisy Logits

Proposed a two-phase training method: train architecturally diversified models individually, then train the ensemble with a diversity-promoting loss. Combined with noise injection at inference, this improves resistance to adversarial attacks while maintaining reasonable accuracy.

Quantifying Deep Learning Model Uncertainty in Conformal Prediction

Proposed a probabilistic approach to quantify model uncertainty derived from conformal prediction sets and provided certified bounds for the computed uncertainty, enabling comparisons between CP-based uncertainty and other uncertainty quantification methods.

Evidence-based Model Predictive Confidence in Deep Classifiers

Developed a method grounded in Dempster–Shafer theory of evidence and Subjective Logic to quantify predictive confidence by separating uncertainty due to ambiguous/conflicting information from uncertainty due to lack of sufficient information.

Ontology Alignment Using Word/Sentence Embedding Methods (M.Sc. Thesis)

Proposed an approach to discover valid mappings between biomedical ontologies using Word2Vec-based representations and semantic similarity (Maximum Matching), combined with retrofitting using WordNet to improve alignment quality.

A Learning-based Approach for Ontology Alignment Using Inductive Logic Programming (ILP)

Developed an ontology mapping method based on Inductive Logic Programming (ILP): interpret OWL ontologies into first-order logic predicates, then apply inductive reasoning with background knowledge to discover hidden rules and propose valid alignments via structural similarities.

Implementing YAD: An Inductive Logic Programming Tool

Implemented an ILP tool for multidimensional and multi-tabular learning using first-order logic representations of positive/negative examples to discover rules and relations.

Reinforcement Learning Approach in Multi-Agent Systems Based on Cooperation (B.Sc. Final Project)

Studied cooperative multi-agent reinforcement learning and proposed an RL-based approach in the problem space (beyond the optimal point), analyzing performance and efficiency tradeoffs.