Vahid Tarokh
Duke University, NC·Engineering
Streams of data sometimes change behavior abruptly: a power line fails, a cyber-attack begins, an epidemic starts, or something unusual enters a video feed. The task is to flag such a change as quickly as possible while rarely crying wolf.
- Fields
Engineering
Mathematics & Statistics
Computer Science
- Topics
quickest change detection
anomaly detection
score matching
- Grants
- 1 known grant, through Apr 2027
What they work on
- Develop the theory of score-based quickest change detection for independent, identically distributed single-stream data under Bayesian, generalized Bayesian and minimax formulations, and establish that performance depends on the Fisher distance between distributions.
- Design robust score-based detection methods that tolerate modeling uncertainty, characterizing performance through Fisher distances between the candidate distributions in an uncertainty class.
- Define a notion of score for dependent data sequences, based on the gradient of the logarithm of conditional densities, and derive change-detection algorithms that are optimal for such sequences.
- Build distributed change-detection algorithms for multiple agents that each know only part of the underlying distributions and communicate only with neighbors, and validate the methods on public anomaly-detection datasets.
More about the research
Classical detection methods assume the probability distributions of the data before and after the change are known or can be estimated, which is unrealistic for the complex, high-dimensional data common in practice.
Why it matters
Faster, more dependable detection of abrupt changes could support anomaly detection in video, line-outage detection in electrical grids, early warning for disease outbreaks, and identification of cyber-attacks. Because score-based methods do not require known distributions, the resulting algorithms could be applied by researchers across fields working with high-dimensional data streams.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Min Xu
Rutgers University-New Brunswick, NJ·Mathematics & Statistics
Develop a general Markovian network model for gene-gene interaction networks that captures the step-by-step growth of a network over evolutionary time and includes existing network models as special cases.
Ting Zhang
University of Georgia, GA·Mathematics & Statistics
Develop change-point tests that stay powerful when the post-change signal varies unpredictably, using a causal representation framework and Gaussian approximation to set statistically sound alarm thresholds.
Jyotishka Datta
Virginia Polytechnic Institute and State University, VA·Mathematics & Statistics·Early-career
Develop Bayesian independent component analysis with heavy-tailed priors and novel latent space representations, unifying feature extraction methods and controlling higher-order moments when separating hidden sources from mixtures.
The school
Duke University: 61 labs on PI Pages, $1.6B a year on research, 417 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.