Existing learned indexes for vector data are predominantly static data structures due to the inability of the internal model to accommodate changes within the underlying partitioning without complete retraining. The objective of this work is to design, implement, and systematically evaluate a dynamic learning index for vector data by applying methods from continual learning to render the model adaptable to the changes.

Specifically, the student will conduct a review of existing literature on continual learning methodologies, critically evaluating and experimentally assessing their feasibility within the context of learned indexing. Subsequently, the student will select appropriate methods and utilize them to design and implement a dynamic learned index for vector data. The index will facilitate the insertion of previously unseen vectors while efficiently adapting its internal structures and, specifically, the model without the necessity of retraining it from scratch. The index will be compared to the static and dynamized versions of Learned Metric Index on a subset of the LAION-5B dataset and the in-distribution dataset agnews-mxbai from VIBE.