Lifelong and continual learning - A survey

Abstract

While traditional machine learning techniques involve one single iteration of the collation, preparation and training of a model with data, this handicaps models by being a ”no updates once deployed” approach. Using this approach not only severely limits the model’s performance with respect to real-world data that might change continually, but also makes several assumptions about the widespread availability of data, that might not always be the case. Data also has the tendency to constantly evolve, with several slight changes over time being adding up to make a large dent in the performance of static models. With these limitations in mind, it is in our best interests to create ”lifelong learners” - models that are both able to learn new tasks, as well as be good at the old ones. This survey paper aims to understand the basics behind the concept of lifelong and continual learning, as well as learn more about the leading ideas in the same field. Additionally, this paper also aims to understand the core concepts and ideology behind successful lifelong learners, and factors that must be taken into account when building one.