Composing without forgetting

In this project, we propose a modular continual learning approach to face the problem of catastrophic forgetting
and transfer in learning from evolving task distributions. Concretely, we propose a model that learns how to select
most relevant modules based on a local decision rule for a given task to form a deep learning model for solving a
given task. In this framework we generalization to unseen but related tasks emerge through the composition of
those modules. Additionally, we exploit self-supervised learning to further boost performance through test-time
self-supervised finetuning (active remembering). This is of vital importance for Element AI to provide reusable
solutions that scale with new data, without the need of learning a new model for every problem and improving the
overall performance.

Faculty Supervisor:

Laurent Charlin

Student:

Partner:

ServiceNow Canada

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

HEC Montréal

Program: