Probabilistic deep learning
New methods and metrics for trustworthy uncertainty quantification in regression — calibration, conditional fit, and full heteroscedasticity for both continuous and count outputs.
Building machine learning systems that are reliable enough to act on under uncertainty — spanning probabilistic deep learning, 3D perception, and real-world decision systems.
I am an AI scientist and researcher with a PhD from Penn State University. I am an Assistant Professor of Computer Science at Brigham Young University (BYU), where I teach courses on Deep Learning, Databases, and Probabilistic Machine Learning. I am currently on leave from BYU to build the future of Agentic Software at Enzy. My research has been published in leading AI conferences, including NeurIPS, ICML, ICLR, EMNLP, KDD, AAAI, IJCAI, WWW, and WACV.
Beyond academia, I have several years of experience collaborating with industry partners to apply AI technologies in real-world settings. My partnerships span both startups — Delicious AI, Enzy Technologies, and NextGen Nuclear — and enterprise organizations like Vivint Smart Home and CandyCo. My work bridges the gap between cutting-edge research and practical AI deployment, helping shape the future of intelligent systems.
My research sits at the intersection of probabilistic deep learning, computer vision, and real-world decision-making systems, with a unifying focus on building machine learning models that are reliable enough to act on under uncertainty.
One thread of my work develops new theory and methods for uncertainty quantification in deep regression — including the Deep Double Poisson Network (ICML 2025), which enables fully heteroscedastic count regression by parameterizing the Double Poisson distribution, and the Conditional Congruence Error (ECAI 2025), a kernel-based metric that diagnoses point-wise probabilistic fit where standard calibration measures like ECE fall short. A second, complementary thread translates these probabilistic foundations into deployed perception and decision systems, especially for physical retail: CountNet3D (WACV 2023) fuses RGB images with LiDAR point clouds to count densely packed, heavily occluded objects on store shelves, and EdgeRec3D (KDD 2024) builds on that perception layer to deliver a real-time, on-edge product assortment recommender driven by a Robust Bayesian Payoff model and spatial clustering of consumer preferences, with live A/B tests demonstrating 27–35% sales lifts.
Taken together, our work argues — and empirically shows — that principled uncertainty modeling and 3D scene understanding are not just academic concerns but the missing ingredients for trustworthy, high-impact ML deployments in the physical world.
New methods and metrics for trustworthy uncertainty quantification in regression — calibration, conditional fit, and full heteroscedasticity for both continuous and count outputs.
Fusing RGB images with LiDAR point clouds to count densely packed, occluded objects in real-world retail and inventory scenes — end-to-end regression where detection-based pipelines fail.
Bayesian reward models and spatial clustering deployed at the edge — turning noisy 3D observations into product-assortment decisions with measurable, double-digit business impact.
Top-tier venues that anchor my research arc — from probabilistic decision-making and 3D perception, through uncertainty quantification, to reliable reasoning and generation in large models.
Selects teacher reasoning trajectories a student can actually learn from: a learnability rate ρ estimates how fast each trajectory drives the student's loss down, and a χ²-regularized closed-form selection policy balances learnability against distributional coverage — roughly 7 points over the strongest baseline on AIME, AMC, GPQA, and MATH-500.
Shows that naive self-correction can amplify hallucinations, then repairs generations at the fact level: independent observation and claim graphs, per-claim risk scores, and evidence routing with a geometric convergence guarantee — across image, image+text, audio, and video inputs, with a CrisisFACTS case study.
A theoretical and practical framework that recasts LLM self-reflection as in-context policy optimization, proving a one-layer linear self-attention transformer can imitate a bandit policy-optimization algorithm and yielding ME-ICPO — top-tier math-reasoning accuracy at modest inference cost.
A neural count-regression model that parameterizes the Double Poisson distribution to give discrete-output networks fully heteroscedastic aleatoric uncertainty for the first time, beating Poisson, Negative Binomial, and Gaussian baselines on accuracy, calibration, and OOD detection.
A kernel-based, point-wise diagnostic for the probabilistic fit of deep regressors that detects misspecification ECE silently misses — and unlike NLL, can be computed at any new input without needing the label.
A real-time product-assortment recommender that uses on-edge 3D vision for sales estimation, a Robust Bayesian Payoff model to handle observation noise, and spatial clustering to personalize across stores — with live A/B tests showing 27–35% sales lifts.
A Bayesian framework for ranking candidate offline policies by their expected return distributions, enabling principled, uncertainty-aware policy selection in the offline RL setting.
A regression-based 3D counting network that fuses RGB and LiDAR via novel "PointBeam" proposals to count densely packed, heavily occluded objects, outperforming detection-based 3D counters on both synthetic and real retail shelves.
A model-based offline RL approach to the combinatorial product-allocation problem, learning a Bayesian world model that captures uncertainty over consumer demand and using it to optimize shelf-level decisions.
A framework that learns economic utility functions with neural networks, bridging classical consumer-choice theory with modern representation learning for recommendation.
Complete publication record across conferences, workshops, and journals. Tier-1 venues are bolded.
My thoughts on a broad range of topics. Key ideas include (but not limited to) AI, probabilistic machine learning, epistemology and Bayesian Inference.
Several years of collaboration deploying AI in the real world — across startups and enterprise. These partnerships ground my research in the operational realities of live data, edge devices, and measurable business outcomes.
3D computer vision and edge ML for physical retail. The technical home of CountNet3D and the EdgeRec3D recommender system, with live deployments delivering double-digit sales lifts.
Applied machine learning collaborations focused on data-driven sales workflows and field-team intelligence.
AI methods for advanced nuclear technologies — bringing probabilistic modeling and decision systems to a high-stakes physical-systems domain.
Long-running collaboration on applied ML at enterprise scale — translating research into capabilities that ship in real consumer products.
Enterprise grade analytics coupled with state-of-the-art probabilistic forecasts of consumer goods demand. Deploying my research into brick and mortar stores.