Developed algorithm based on convolutional neural networks and other computer
vision methods to detect rare GNSS interference events in geolocation (AIS, ADS-B)
data as a key deliverable for a Department of Defense contract.
Trained convolutional neural networks on generated synthetic data using PyTorch
on AWS EC2 instances.
Variational inference for robust Gaussian process regression
We developed a scalable approximate inference algorithm for robust Gaussian process regression with contaminated normal noise. Initial experiments with simulated and actual data are promising.
We introduced the notion of ‘mechanistic heterogeneity’ in Mendelian Randomization, developed a latent mixture method to model it, and used it to estimate heterogeneous causal effects between HDL cholesterol and coronary heart disease.