Posts by Collection

portfolio

Portfolio item number 1

Short description of portfolio item number 1
Read more

Portfolio item number 2

Short description of portfolio item number 2
Read more

publications

Real-Time Particle Collision Filtering with Unsupervised Learning on Custom FPGA Hardware

Authors: Tanish Kumar, Zakk Heile, Nika Kiladze, Zesen Zhuang, Ashutosh Kotwal

Duke Research Symposium, 2024

At CERN’s Large Hadron Collider, physicists recreate the conditions of the early universe to discover the constituents of dark matter by experimentally observing the products of proton-proton collisions. However, with protons colliding every 25 nanoseconds, computationally processing all the detector data from every collision is an impractical task. Our work is focused on designing and implementing a filtering algorithm that processes collision data, within the significant time constraint, to assess whether each collision warrants further study. Read more

Context-Aware Filtering of Unstructured Radiology Reports by Anatomical Region

Authors: Zakk Heile, Pranav Manjunath, Brian Lerner, Samuel Berchuck, Monica Agrawal, Timothy W. Dunn

Machine Learning for Health (ML4H), 2025

Radiology reports contain essential clinical information but are often stored as unstructured free text. In trauma settings, multiple imaging studies (e.g., CT head, facial bones, and cervical spine) may be bundled into a single report that consolidates findings from all examinations into one jointly written document. As a result, individual sentences may reference ambiguous or overlapping anatomy (e.g., “there is a fracture”), making sentence-level anatomical classification essential for downstream tasks. Read more

See Paper

Discovering AI Applications for Traumatic Brain Injury Care

Authors: AI4TBI Bass Connections Team

Bass Connections Research Showcase, Duke University, 2025

We present a mixed-methods investigation of traumatic brain injury (TBI) care at Duke University, aimed at identifying opportunities for responsible and clinically aligned AI decision support. The project combines rapid qualitative analysis of clinician interviews and shadowing across the TBI care pathway with large-scale curation and quality assurance of multimodal electronic health record data. Findings highlight key documentation challenges, sources of clinical variability, and stakeholder-guided requirements for AI deployment, alongside the construction of a curated TBI dataset to support future modeling and benchmarking. Read more

Scalable Rigid-Invariant Distance for Shape Matching and Alignment

Authors: Zakk Heile, Peilin He, Jayson Tran, Ruiling Wang, Shrikant Chand

NeurIPS Workshop for Imageomics: Discovering Biological Knowledge from Images Using AI, 2025

Comparing probability distributions derived from biological images requires distances that are geometrically grounded and invariant to orientation. Classical optimal transport (OT) distances are sensitive to rotations, while Gromov–Wasserstein (GW) offers invariance but is computationally prohibitive for large datasets. Read more

See Paper

Efficient Rashomon Set Approximation for Decision Trees

Authors: Zakk Heile, Varun Babbar, Hayden McTavish, Cynthia Rudin

NeurIPS Workshop on Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making (MLxOR), 2025

Standard machine learning pipelines often admit many near-optimal models. These Rashomon sets pose both challenges and opportunities for uncertainty-aware and robust decision making: they enable incorporation of domain knowledge and user preferences that are difficult to encode directly in an objective, and they quantify diversity among plausible models and predictions for a given dataset and loss. Read more

See Paper

Rigid-Invariant Sliced Wasserstein via Independent Embeddings

Authors: Zakk Heile, Jayson Tran

Joint Mathematics Meetings (JMM), 2026

Comparing probability measures whose supports are related by an unknown rigid transformation is a fundamental challenge in geometric data analysis, arising in shape matching and machine learning. Classical optimal transport (OT) distances, including Wasserstein and sliced Wasserstein, are sensitive to rotations and reflections, while Gromov–Wasserstein (GW) is invariant to isometries but computationally prohibitive for large datasets. Read more

Rigid-Invariant Sliced Wasserstein via Independent Embeddings

Authors: Zakk Heile, Peilin He, Jayson Tran, Alice Wang, Shrikant Chand

ICLR GRaM Proceedings, 2026

Comparing probability measures when their supports are related by an unknown rigid transformation is an important challenge in geometric data analysis, arising in shape matching and machine learning. Classical optimal transport (OT) distances, including Wasserstein and sliced Wasserstein, are sensitive to rotations and reflections, while Gromov–Wasserstein (GW) is invariant to isometries but computationally prohibitive for large datasets. Read more

See Paper

Align Representations, Not Points: Efficient Rigid Invariant Transport Distance

Authors: Zakk Heile, Peilin He, Jayson Tran, Alice Wang, Shrikant Chand

ICLR Re-Align, 2026

Comparing probability measures modulo unknown rigid transformations is a central challenge in geometric data analysis. Classical optimal transport (OT) distances, including Wasserstein and sliced Wasserstein, are sensitive to rotations and reflections, whereas Gromov-Wasserstein (GW) and Procrustes-Wasserstein (PW) distances are invariant to isometries but computationally prohibitive for large datasets. Read more

See Paper