Xiao Fu

What they work on

1. Build a functional equation-based framework that proves when nonlinear factor analysis models, which extend NMF and independent component analysis to unknown nonlinear distortions, can be uniquely identified without supervision.
2. Characterize nonlinear factor analysis under limited, noisy data and develop algorithms with performance guarantees, evaluating them on hyperspectral imaging, biosensor signal processing and unsupervised machine learning.
3. Develop a coupled nonnegative matrix factorization framework that links the Dawid-Skene annotator model to nonnegative factor analysis theory, to label massive datasets from noisy, incomplete crowdsourced annotations with provable guarantees.
4. Design situation-aware crowdsourcing methods that exploit data structure and annotator dependence, along with scalable stochastic optimization, and evaluate them on real-world labeling applications.
5. Develop latent neural factor analysis (LaNFAC) models, which combine tensor decomposition with neural generative models, to reconstruct radio maps of complex, shadowed and dynamic environments from a few quantized measurement bits.
6. Create provable spectrum cartography using untrained LaNFAC models that need no training data, and validate it with simulators and software-defined radio experiments on real data.
7. Establish provable fusion of unregistered hyperspectral and multispectral images by combining spectral unmixing with adversarial learning through distribution matching in a latent spatial domain.
8. Extend unregistered spectral image fusion to large, unknown nonrigid deformations and design stable, efficient optimization algorithms, validated on semi-realistic and real-world datasets.
9. Develop flow matching learning objectives and transformation paths, which learn velocity fields between data domains, along with conditions that guarantee unique, content-preserving translation across unpaired domains.
10. Establish theory and efficient algorithms for controllable multi-domain generation with flow matching, so that desired content or style is reliably preserved under limited supervision.

More about the research

Why it matters

Methods with guarantees could make data analysis more reliable when hidden distortions or misaligned data are involved. Better fusion of unregistered spectral images could aid agriculture, oceanography, forest monitoring and mineral discovery. Radio maps built from a few measurement bits could help wireless networks manage spectrum and avoid interference. Theory-backed crowdsourcing could reduce annotator workload when labeling data for AI. Identifiable flow matching could make AI translation and generation less prone to hallucinations in science and healthcare.

Known grants

From NIH and NSF research grants. Labs often have other funding we can't see.

Closest labs

The school

Oregon State University: 29 labs on PI Pages, $372M a year on research, 238 research doctorates a year.

More about Oregon State University

Reach out

A start for your email. Fill in the brackets.