Projectsedit

Research projectsedit

AI and networksedit

AI and Networks is a current research focus. It includes AI for Networks, Networks for AI, data pruning for decentralized learning, communication-aware evaluation, reliability across data silos, and distributed computation of Wasserstein-based reference distributions.

Distributed Wasserstein barycentersedit

Distributed Wasserstein barycenter is part of this work on AI and networks. It asks how multiple parties can compute or approximate a shared distributional reference from local empirical measures, with applications to collaborative evaluation, sample scoring, and synthetic-data verification.

Machine unlearningedit

Machine Unlearning includes both approximate certified unlearning for differentiable models and exact or efficient unlearning for tree ensembles. Related work includes Hessian-Free Online Certified Unlearning, Beyond Binary Erasure: Soft-Weighted Unlearning for Fairness and Robustness, and DynFrs: An Efficient Framework for Machine Unlearning in Random Forest.

Collaborative evaluationedit

Collaborative Evaluation studies verification without raw-data exchange. It is used in the ICML 2026 model-collapse work to replace a single low-resource, biased verifier with multi-party Wasserstein-geometry proxies.

Synthetic dataedit

Synthetic Data asks when generated data can safely replace or augment real data, and when recursive training amplifies bias or erodes diversity. The current emphasis is low-resource communities, where fragmented real-data coverage makes local filtering more likely to prune valid tail modes. A central paper is When Sample Selection Bias Precipitates Model Collapse.

Trustworthy LLM systemsedit

LLM Reliability examines whether language models infer relations without adequate evidence. The NeurIPS 2026 paper Illusory Pattern Perception Drives Spurious Inference in Large Language Models compares LLM and human illusory pattern perception across three psychological tasks.