Call for papers
Current supervised machine learning approaches have shown extraordinary capabilities in solving complex tasks, but the resulting models still lack the typical human skill of being able to sequentially learn and remember how to perform multiple tasks, or to deal with changing data distributions. Continual learning approaches aim at tackling these limitations, by devising techniques that allow models to retain or refresh the knowledge of previously-learned tasks, thus preventing “catastrophic forgetting”.
This family of approaches has recently attracted a lot of attention, but experimental evaluations have been mostly limited to constrained and artificial scenarios (e.g., presence of hard boundaries between tasks; assumption that classes cannot reappear in later tasks, lack of correlation between consecutive tasks, etc.). This can undermine the applicability of continual learning to real-world problems: for example, robotics would benefit from autonomous agents that are able to adapt to new terrains while still being able to navigate on the ones they already know; in medicine, privacy concerns may limit the availability of historical data, and domain shifts (e.g., imagery acquired by new equipment or disease progress) would require that models be able to process both data from the old distribution, now unavailable for training, and the new one.
To tackle the challenges posed by such settings, common pitfalls and possible amendments of the classical scenarios have to be investigated in the next few years. Therefore, this workshop aims to promote and attract research on less studied aspects of continual learning, from new benchmarks to evaluation protocols, thus integrating the current methodological research efforts with a more practical and application-oriented perspective.
The workshop aims to attract novel and original contributions exploring the intersection of continual learning and real-world applications. Expected submissions should cover, but are not limited to, the following topics:
- investigations and proposals of metrics to better qualify the performance of continual learning approaches in the context of realistic problems;
- novel continual learning benchmarks that go beyond the classical settings, focusing on one or more key aspects that characterize applications;
- case studies of deployment and integration of continual learning approaches to real-world problems;
- methodological contributions that focus on key aspects of realistic applications.
Workshop proceedings will be published by Springer LNCS.