Learning from Small Data Sets and Incremental Learning

Description of the research area

Recent successes in Deep Learning relies on large annotated data sets. This cluster focuses on methods that make machines able to learn effectively from less (annotated) data. This could enable agents to learn incrementally, in a life long learning experience. We cover application as well as theory.

The following keywords show some problems and techniques of interest:

• Semi-supervised: Using a smaller set of annotated data together with a bulk of unannotated data.
• Active learning: Letting the agent point out the most relevant data for annotation.
• Data-augmentation and synthetic data to scale up smaller datasets.
• Transfer learning to repurpose a model for a different task.
• Few-shot learning: Learning a completely new task from only a few samples.
• Meta-learning: Learning to learn.
• Incremental learning: Ability to learn a new task without forgetting the earlier tasks.

Cluster Activities

• A meetup per month (online) to share experience and information.
• Physical meetup at the yearly Winter Conference (in January).
• A slack channel under Wasp-Sweden, for ad hoc discussions.

Note: The detailed activity log is kept in an internal document. Contact cluster leader for access.

Contact all Members

Please use responsibly!

Mailing List: CTC_SmallIncremental@wasp-sweden.se

Other Information

The cluster was set up in March 2020 and consists of ~20 PhD students currently signed up. Contact the cluster leader if you want to join or leave.

Even though this is one of the larger clusters, the majority of the members are passive. If you can share insights for this area, you’re very welcome to give a presentation. If you just want to keep an eye on what happens in this area within WASP, we welcome questions too.