My research in plots

LoFi v.s. FC-EKF

Our Low-rank Extended Kalman filter (LoFi) method for distribution shift compared the to full-covariance extended Kalman filter (FC-EKF). We consider a neural network with 151 parameters. We approximate the high-dimensional distribution of the parameters with a 10-dimensional low-rank space.

posterior-predictive-comparison (1).mp4

In-between uncertainty for a Bayesian Neural network

Posterior predictive distribution of a Bayesian neural network: we sequentially add a new datapoint and draw samples from the posterior using Hamiltonian Monte-Carlo (HMC). The underlying data distribution was generated so that no data is seen at the outer tails and in the middle.

animation-hmc (3).mp4

Online multinomial classification using the exponential-family extended Kalman filter

Training a Bayesian neural network on an online multinomial problem