Hao Yan

Associate Professor in School of Computing and Augmented Intelligence

Arizona State University

Biography

Research Interest

Develop reinforcement learning, generative modeling, and probabilistic machine learning methods for high-dimensional, spatio-temporal, and sequential data (images, profiles, signals, event sequences) in industrial, energy, transportation, and healthcare systems, with a focus on anomaly detection, prognostics, and decision making. Recent directions include distributional and flow-based reinforcement learning, flow-matching and diffusion models for anomaly detection and surrogate modeling, causal structure (DAG) learning, Bayesian and physics-informed learning, and tensor methods.

Download my resumé .

Interests

Generative Models
Reinforcement Learning
Anomaly Detection

Education

Ph.D. in Industrial Engineering, 2017
Georgia Institute of Technology
M.S. in Computational Science and Engineering, 2016
Georgia Institute of Technology
M.S. in Statistics, 2015
Georgia Institute of Technology
B.S. in Economics, 2011
Peking University
B.S. in Physics, 2011
Peking University

Recent News

All news»

Sep 2026, two papers were accepted to NeurIPS 2026: Flow Mismatching: Unsupervised Anomaly Detection via Velocity Discrepancies in Flow Matching Models (Shengzhe Chen, Mehrdad Moradi, Kamran Paynabar) and Structurally Separated DAG Learning with Multi-Scale Normalized Closure (Shuai Feng). Congratulations to Shengzhe and Shuai!
Jul 2026, our paper Adaptive Sampling for Automated Post-Disaster Rapid Damage Assessment via Level-Set Cost-Aware Bayesian Optimization (Boyang Xu, Mostafa Reisi Gahrooei, Mohammad Ilbeigi) was accepted to SDM 2026 (SIAM International Conference on Data Mining) in Salt Lake City.
Jun 2026, our paper D-Convexity: A Unified Differentiable Convex Shape Prior via Quasi-Concavity for Data-driven Image Segmentation (with Shengzhe Chen) appeared at IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 as a Highlight (top 2%). Project page · arXiv · Code · Virtual poster.
May 2026, our paper Path-Coupled Bellman Flows for Distributional Reinforcement Learning was accepted to International Conference on Machine Learning (ICML) 2026 as a regular presentation (authors: Boyang Xu, Qing Zou, Siqin Yang, Hao Yan). Flow matching with path coupling for distributional RL; improves stability on offline benchmarks including OGBench and D4RL.
May 2026, our paper Aleatoric-Aware Posterior Regularized Bayesian Neural Networks for Incorporating Domain Knowledge: Addressing Epistemic and Aleatoric Uncertainties (Jiayu Huang, with Jing Li, Frank Setzer, Fleming Lure, Feng Ju) was accepted to INFORMS Journal on Data Science. Congratulations to Jiayu!

Featured Publications

Shengzhe Chen, Mehrdad Moradi, Kamran Paynabar, Hao Yan

September, 2026 Advances in Neural Information Processing Systems (NeurIPS)

Flow Mismatching: Unsupervised Anomaly Detection via Velocity Discrepancies in Flow Matching Models

Unsupervised anomaly detection from the mismatch between a flow matching model’s learned velocity and the direct velocity toward a test image, giving pixel-wise heatmaps and image-level scores.

Shengzhe Chen, Hao Yan

January, 2026 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

D-Convexity: A Unified Differentiable Convex Shape Prior via Quasi-Concavity for Data-driven Image Segmentation

A unified, threshold-free differentiable convex-shape prior based on quasi-concavity of the segmentation mask function. Zero-, first-, and second-order characterizations yield a midpoint convexification algorithm and compact convolutional losses that integrate seamlessly with modern segmentation networks via the proposed Convex Gradient Projection Module (CGPM). Accepted to CVPR 2026 as a Highlight.

Boyang Xu, Qing Zou, Siqin Yang, Hao Yan

January, 2026 Proceedings of the International Conference on Machine Learning (ICML)

Path-Coupled Bellman Flows for Distributional Reinforcement Learning

Flow-matching distributional RL with source-consistent Bellman-coupled paths and shared-noise coupling between current and successor return flows. A lambda-parameterized control variate reduces bootstrap variance while preserving Gaussian source and Bellman endpoint geometry. Strong distributional fidelity on toy MRPs and competitive offline RL on 38 OGBench and D4RL tasks. Accepted to ICML 2026 as a regular presentation.