We find that early-trunctated conjugate gradients tends to underﬁt while random Fourier features tends to overﬁt in Gaussian processes learning.
We address these issues using randomized truncation estimators that eliminate bias in exchange for increased variance.
We propose HIP-GP, an inter-domain GP inference method that scales to millions of inducing points.
HIP-GP relies on gridded inducing points and stationary kernel assumptions, and is suitable for low-dimensional problems.
We propose a pipeline for body dynamics inference from video:
we use a convolutional network to track joint positions, and embed these
as the joints of a linked robotic manipulator; we develop a probabilistic physical
model whose states specify second-order rigid-body dynamics and design a distributed nested SMC inference algorithm.