Sequoia: A Software Framework to Unify Continual Learning Research

  • 2021-08-02 16:07:21
  • Fabrice Normandin, Florian Golemo, Oleksiy Ostapenko, Pau Rodriguez, Matthew D Riemer, Julio Hurtado, Khimya Khetarpal1, Dominic Zhao, Ryan Lindeborg, Thimothée Lesort, Laurent Charlin, Irina Rish, Massimo Caccia
  • 19

Abstract

The field of Continual Learning (CL) seeks to develop algorithms thataccumulate knowledge and skills over time through interaction withnon-stationary environments and data distributions. Measuring progress in CLcan be difficult because a plethora of evaluation procedures (\emph{settings})and algorithmic solutions (\emph{methods}) have emerged, each with their ownpotentially disjoint set of assumptions about the CL problem. In this work, weview each setting as a set of \emph{assumptions}. We then create a tree-shapedhierarchy of the research settings in CL, in which more general settings becomethe parents of those with more restrictive assumptions. This makes it possibleto use inheritance to share and reuse research, as developing a method for agiven setting also makes it directly applicable onto any of its children. Weinstantiate this idea as a publicly available software framework called\emph{Sequoia}, which features a variety of settings from both the ContinualSupervised Learning (CSL) and Continual Reinforcement Learning (CRL) domains.Sequoia also includes a growing suite of methods which are easy to extend andcustomize, in addition to more specialized methods from third-party libraries.We hope that this new paradigm and its first implementation can serve as afoundation for the unification and acceleration of research in CL. You can helpus grow the tree by visiting \url{www.github.com/lebrice/Sequoia}.

 

Quick Read (beta)

loading the full paper ...