Jiafang Song

PhD Candidate, Biostatistics
Johns Hopkins University

jsong100@jhu.edu

I work on Bayesian inference, machine learning, variational inference, normalizing flows, and spatial statistics.

Ph.D. Research

  • Cutting Feedback for Uncertainty Propagation using Neural Networks
    Designed neural network-based method for efficient uncertainty propagation between upstream and downstream analyses without reverse back. Approximated cut-posteriors using conditional normalizing flow. More information on arxiv.
    NeVI-Cut: A python package for modular Bayesian inference, directly utilizes posterior samples from upstream and triply stochastic scalable algorithm is devised to implement the method.

  • Fast Variational Inference for NNGP
    Designed a fast variational inference method for posterior approximation in high-dimensional linear mixed-effects models. More information on arxiv.
    spVarBayes: An R package for fast variational Bayesian inference in large-scale geospatial regression using NNGP. It leverages closed-form gradient updates and linear response corrections to achieve calibrated uncertainty at reduced computational cost. A tutorial is available.

Collaborations with Infectious Disease Dynamics

  • Disease Mapping
    Extended the directed acyclic graph autoregressive (DAGAR) model to accommodate multi-source, misaligned, and censored data, and applied it to quantify spatial and spatio-temporal patterns in real-world datasets.

Master’s Thesis

  • Spatial Modeling of Point Source Influence
    Developed a hierarchical Bayesian model with MCMC sampling to estimate both magnitude and spatial extent of influence from a point source, incorporating a Gaussian predictive process for large-scale data. Paper is available at JABES.

Master’s Research

  • Measurement Error and Causal Inference
    Derived exact and asymptotic properties of regression calibration for causal effect estimation in truncated and censored survival data.
    Developed EM algorithms for Cox models with covariate measurement error, and conducted simulation studies benchmarking against conventional regression calibration methods.