May 2022 Summaries

Reverse image search is an application of vector search or approximate nearest neighbor search. In this process, indexes are built to accelerate the search on large datasets. This article discusses how to visualize reverse image search with Feder using the example of IVF_FLAT index. The IVF_FLAT index divides vectors in the vector space into different clusters based on vector distance. During a vector similarity search, users need to provide a target vector and the configuration of search parameters. Feder then visualizes the whole search process. In this use case, we use VOC 2012 dataset with an nlist of 256 to build an IVF_FLAT index. The system first calculates the distance between the target vector and the centroid of each cluster to find the nearest clusters. Then it compares the distance between the target vector and all vectors in the nprobe clusters for a fine search. Feder provides two visualization modes for the fine search process, one based on cluster and vector distance, and the other is the projection for dimension reduction mode. The value of index building parameters influences how the vector space is divided, and the nprobe parameter can be used to achieve tradeoff between search efficiency and accuracy.

Feder is a tool that enables users to visualize approximate nearest neighbor search (ANNS) algorithms, specifically indexes like IVF_FLAT and HNSW. It helps users understand the structure of different indexes, how data are organized using each type of index, and how parameter configuration influences the indexing structure. Feder currently supports the HNSW from hnswlib but plans to support more indexes in the future. The tool is built with JavaScript and Python, allowing users to visualize index structures and search processes under IPython Notebook or as an HTML file for web service use.