Collaborative Research: ATD: Robust, Accurate and Efficient Graph-Structured RNN for Spatio-Temporal Forecasting and Anomaly Detection

The project aims to develop robust, efficient, and transferrable deep learning algorithms for prediction and anomaly detection in human spatio-temporal dynamics. This will be a fundamental step in providing reliable and speedy decision support for mitigating infectious diseases and countering threats in a time varying and spatially complex environment.

The project shall advance recent computational tools (deep neural networks) in adversarial conditions and on resource limited (low cost, low energy) platform, thereby contribute to information technology in adversarial learning, mobile computing and effective decision making. A broad range of applications include threat detection and prediction for traffic and public transportation networks, security and privacy critical data analysis and prediction, threat detection and error correction for hydraulic, electrical and nuclear power systems.

The approaches to be used involve novel techniques in high dimensional non-smooth non-convex optimization and graph representation. Specifically, the project shall study (1) multi-scale graph-structured recurrent neural networks for spatio-temporal data modeling, prediction and anomaly detection; (2) adversarially robust, accurate, and transferable deep learning algorithms based on advection-diffusion equations; (3) efficient quantization algorithms under adversarial conditions to reduce the latency of deep networks.

The projects shall train a diverse body of graduate and undergraduate students at the Irvine and Los Angeles campuses of University of California through collaborative education and research activities in applied mathematics, computer science, data science and social science.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.