In the last few years we have witnessed a renewed and steadily growing interest in the AI community towards algorithms that can learn continuously from high-dimensional data. At ContinualAI we work to create a distributed and inclusive research lab on Continual Learning, where anyone can contribute and learn more about this fascinating topic, while producing cutting edge research results.

We believe in open science at every level: from the decision process to the development and the release of the research products. While in recent years the AI community has started open-sourcing the final research products (e.g. paper and software), research is still conducted within closed doors and small research labs. At ContinualAI we believe in a more inclusive approach where research is conducted openly at every stage with the possibility of huge benefits for the community.

In particular, current research projects at ContinualAI include:

  • Avalanche: A comprehensive framework for Continual Learning Research. It aims at unifying a set of popular CL baselines, environments and benchmarks to help algorithm prototyping and experiment, with flexibility, reproducibility, efficiency and maintainability in mind. Avalanche will be based on three main modules: datasets/environments, CL baselines and evaluation metrics/protocols. See the projects page for more details.

  • Short-Science Summary: Even considering only the topic of Continual Learning, keeping up with the huge amount of papers published today can be very difficult. This is why, in this project, we plan to contribute to the awesome short-science with short descriptions of CL papers. See the projects page for more details.

  • CL Paper Database: Waiting for better AI tools for papers recommendation the ContinualAI community is maintaining a database of CL papers which we plan to release soon. It would be very rich of meta-data so that we can better navigate the incredible number of papers published each year (query example: give me the papers employing rehearsal and evaluated on CORe50). See the projects page for more details.