UNCW Data Science Seminar

The Data Science/AI Seminar is interdisciplinary. All faculty and students are welcome. Email DataScienceUG@uncw.edu if you are interested in presenting or would like to join the seminar mailing list.

Note: The time and location may vary.

Fall 2026 Schedule

Abstract

Bilipschitz embeddings for invariant machine learning - D. Mixon

Machine learning algorithms are typically designed for Euclidean data, but many natural datasets come with symmetries: a group G of isometries acts on a Euclidean space V, and points in the same orbit represent the same object. That means the true data space is not V, but the orbit space V/G. Invariant machine learning represents this quotient by a G-invariant feature map into Euclidean space. For robustness, especially against adversarial examples, this feature map should be bilipschitz with respect to the quotient metric. Sadly, vanilla polynomial invariants fail to be bilipschitz, so we need to move beyond classical invariant theory. In this talk, we present low-distortion embeddings in a variety of settings, and we conclude with applications and several open problems.

AI, Your Career, and the Skills That Still Matter - B. Stensvad

UNCW is pleased to welcome Bryan Stensvad for a guest talk on applied AI in industry. Bryan leads advanced analytics and AI initiatives that directly move the needle for a global technology company — from dynamic pricing recommendation engines and sales productivity platforms to fraud detection systems — work that has driven measurable revenue growth, margin expansion, and operational efficiency at scale. With a career spanning strategy consulting, corporate strategy, and hands-on AI implementation, Bryan brings a rare combination of technical depth and business acumen, offering students a firsthand look at how AI translates into real bottom-line impact inside a Fortune Global 500 company.

40-minute talk + 20-minute open Q&A
Ask Bryan directly about career strategy, industry trends, and breaking into AI-driven roles

Unveiling the Invisible: PM2.5 at NC Schools & Its Academic Shadows - R. Carroll

Fine particulate matter (PM2.5) is a major environmental health concern, particularly for children, yet estimating exposure at school locations remains challenging due to limited monitoring coverage. This study estimates PM2.5 concentrations at North Carolina public schools by combining high-quality but sparse EPA monitoring data with dense, low-cost PurpleAir sensor data. Two interpolation approaches—Kriging and inverse distance weighting (IDW)—were compared using leave-one-out cross-validation. Results show that Kriging consistently outperformed IDW, and that incorporating PurpleAir data improved predictive accuracy, especially in areas lacking EPA monitors. School-level estimates revealed higher PM2.5 concentrations in suburban areas relative to urban and rural settings. Linking these estimates to educational outcomes, higher PM2.5 levels were associated with lower 8th-grade reading scores, though effects were modest. Overall, the study demonstrates that integrating low-cost sensors with traditional networks enhances air pollution exposure assessment and suggests potential implications of air quality for student academic performance.

Leveraging Satellite Data and Geospatial AI for Agricultural Monitoring and Decision Support - Z. Yang

Satellite remote sensing provides repeated observations of ecosystem dynamics across different wavelengths and at various spatial and temporal resolutions. The increasing availability and capability of satellite observations have created new opportunities for remotely monitoring agricultural systems across broad spatial scales throughout the growing season. In this talk, we focus on how Geospatial AI can help integrate multi-source satellite data with environmental observations to enable timely agricultural monitoring under varying growing conditions. Specifically, we discuss a variety of applications ranging from in-season crop type classification, phenology monitoring, to yield prediction. Building on these applications, we are developing integrated approaches to assess flooding impacts on agriculture in the Inner Coastal Plain of North Carolina, with the broader goal of providing decision support for agricultural resilience planning.

Agentic Mathematics with Open Models on Self-Hosted GPUs - T. Whitehouse

Math is being disrupted by the "foundational models" used by service providers like Anthropic and OpenAI. Unlimited budgets give them millions of dollars in GPU hours to run agents and then and run theorem checkers, iteratively.
This gives internet scale businesses leverage to prove or disprove outstanding conjectures, but it's unclear if it's a good model for mathematics. First, it is not affordable. Second, the results are often difficult to understand no matter what the theorem checkers say. And finally, it seems to ignore the overall intent of mathematics, i.e. human understanding.
This talk offers a small alternative to the proprietary, foundational model approach by using open agent harnesses and models to build understanding in a progressive way similar to traditional research yet at a much higher speed. We will go into details for how to get the harness and models set up on GPUs, as well as ideas for developing customizable "tooling" aimed at mathematical research.

Understanding NeuroEvolution of Parameterized Quantum Circuits Using Agentic Systems - T. Desell

NeuroEvolution is the process of utilizing evolutionary algorithms to automate the design and training of neural networks. Our recent work developing the Evolutionary eXploration of Augmenting Quantum Circuits (EXAQC) framework has shown that NeuroEvolution strategies can also be applied to quantum machine learning. In particular, EXAQC has shown it is possible to design and train parameterized quantum circuits which can act as more parameter efficient components within larger neural models across classification, reinforcement learning and vision tasks. However, NeuroEvolution algorithms like EXAQC can train and evaluate tens of thousands of models in a single run and these runs further involve complex population and inheritance dynamics which can be challenging to understand at such scales. In addition to presenting EXAQC and current results, this talk will focus on how we have designed EXAQC in such a way that its NeuroEvolution runs can be easily stored and accessed by a Model Context Protocol (MCP) server, allowing agents to examine ongoing and historical results to provide quicker and deeper insights into their performance, bottlenecks and even find bugs. The talk will put a particular focus on generic takeaways for how to effectively design similar frameworks to take advantage of modern agentic systems.

Ladder-Style Value Transformers for LLM Reinforcement Learning - G. Keith

Critic-free methods such as GRPO have become the default for reinforcement learning on language models, in part because learned critics have been either expensive (a value model at LLM scale) or weak (a value head on the final hidden state, which competes with the next-token objective). GRPO, however, assigns a single sequence-level advantage per rollout, which ignores the temporal structure of multi-turn, partially observable tasks. Its group normalization can also amplify small differences in dense partial rewards over sparser terminal rewards.
I present a lightweight critic inspired by ladder side-turning. A small transformer (about 25.8M parameters, 0.6% of the total) reads intermediate activations from several layers of the frozen base model through stop-gradient gated projections. It has its own attention layers, so it can integrate history independently of the policy. The policy is trained with LoRA, and the critic produces token-level values trained with bootstrapped GAE.
I fine-tune Qwen3-4B-Instruct-2507 on Wordle, modeled as a token-level POMDP. PPO with this critic raises the solve rate from 1% to 99% in 11 hours on a single RTX 5090, with no SFT warmup and no KL penalty, and it adds only about 6% update overhead compared with GRPO. Cold-start GRPO stays near a 1% solve rate, stuck maximizing partial credit. The value transformer outperforms a final-latent critic in solve rate, stability across seeds, and explained variance (0.94 vs 0.82). Bootstrapped returns outperform Monte Carlo returns, which diverged in half of the runs. These results suggest that cheap, accurate value estimation makes actor-critic methods competitive again for multi-step LLM training

Machines and Minds: What Can They Teach Us About Each Other? - J. Keith

This talk examines four questions comparing biological brains and modern AI systems: whether AI develops brain-like codes for vision and language, how brains and AI systems learn without erasing old knowledge, whether groups of AI agents behave like human communities, and whether AIs can accurately report their internal states. Across all four, striking parallels and revealing differences emerge. An AI can match human behavior, or even brain activity, while computing in a completely different way. The most valuable results come when AI makes predictions that researchers can test directly in real brains and people. One example is an AI built to play Atari games that correctly predicts how dopamine neurons respond to reward. AI’s real promise for psychology isn't to imitate minds, but to help us understand them.