uber-research / vargp

Variational Auto-Regressive Gaussian Processes for Continual Learning

Continual Learning with GPs

This repository hosts the code for Variational Auto-Regressive Gaussian Processes for Continual Learning (VAR-GPs) by Sanyam Kapoor, Theofanis Karaletsos, Thang D. Bui.

What can VAR-GPs do?

By first, only training on classes 0/1 (Task 0), and next, only training on classes 2/3 (Task 1), we show below the posterior predictive density surface for a classifier on a synthetic classification dataset with four classes. The marks in red are the training data for each class. Qualitatively, brighter regions represent a higher predictive probability that the classifier assigns to the corresponding class in each figure.

VAR-GPs preserve information from Task 0 even after training on Task 1, i.e. prevent catastrophic forgetting (seen as bright regions preserved around data from Task 0).	VCL (our main baseline), on the other hand, is overconfident in its predictions, often in the wrong regions, and is structurally less stable (seen as large variations in predictive probabilities between Task 0 and Task 1)

Setup

From the root of the directory

conda env create -f environment.yml

This creates the environment from environment.yml.

In addition, ensure availability of the this directory in PYTHONPATH for module imports. As an example in Bash,

export PYTHONPATH="$(pwd):${PYTHONPATH}"

Experiments

All experiment scripts utilize Fire and the arguments can directly be converted to CLI arguments. Appropriate functions are mentioned to look for CLI arguments to change. The default arguments are enough to reproduce results in the paper.

Toy Dataset

CLI Arguments: See toy method.

python experiments/vargp.py toy

Split MNIST

CLI Arguments: See split_mnist method.

python experiments/vargp.py s-mnist

Permuted MNIST

CLI Arguments: See permuted_mnist method.

python experiments/vargp.py p-mnist

Checkpoints and Graphs

The checkpoints files are available under notesbooks/results. Use Git LFS to make sure these are pulled alongside the repository.

TIP: Use git lfs fetch if Git LFS was installed after the first clone.

All graphs in the paper can now be generated via code in the notebooks.

License

Apache 2.0

For research purpose only. Support and/or new releases may be limited.