Skip to content

Bayesian Meta-Learning Through Variational Gaussian Processes

Notifications You must be signed in to change notification settings

vivekmyers/vmgp

Repository files navigation

vmgp

This is the repository of Vivek Myers and Nikhil Sardana for our CS 330 final project, Bayesian Meta-Learning Through Variational Gaussian Processes.

Paper: https://arxiv.org/abs/2110.11044

summary of results

Above: Predictions (bottom row) and posterior predictive distribution densities (for the single red testing point, top row) from each model on a 5·floor(z) latent function task. VMGP is able to predict a multimodal mixture of point distributions to model its uncertainty, outperforming the other, less expressive methods.

prediction examples

Above: Sinusoid Regression Examples. Columns, from left: EMAML, AlPaCA, DKT, VMGP (ours). Each row shows an example test task from a different variant of the regression environment. Rows, from top: standard sinusoids, high frequency sinusoids, out of range testing, and tangent functions.

About

Bayesian Meta-Learning Through Variational Gaussian Processes

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published