PyTorch implementation + improvement of Variational Continual Learning (Ngyuen et al.) using Laplacian variational distribution.
All modules and experiments can be found in the laplacian-VCL.ipynb notebook (which may render properly in your GitHub if one is lucky).
We invite those curious to try out different hyperparameter configurations and see the resulting changes.
Read our full report on Overleaf.
Utilizing the same experimental setup as VCL, we find that Laplacian variation leads to superior performance at the end of training:
as well as throughout the learning process:
