Tasks of the Implementation Challenge and Demo Track

Test Data and Queries: Approximately 100 million CLIP descriptors extracted from the LAION database Similarity between two objects is measured by their dot product. The similarity function is a dot product The goal is to evaluate 30 nearest neighbours queries using 10K query objects described below

Task 1: Unrestricted Indexing

In this task, system solutions will have access to all resources on our testing computer to build their indexing solutions. The goal is to achieve the highest search performance within the given constraints.

• Wall clock time for index construction: 12 hours.

• Ranking: Highest throughput/fastest search time with an average recall of at least 0.8.

• Search performance will be evaluated using a built index (in a single configuration) and various query executions using up to 30 different search hyperparameters.

• Saving the index is not required for running the search.

Task 2: Memory-Constrained Indexing with Reranking

This task challenges participants to develop memory-efficient indexing solutions with reranking capabilities. Each solution will be run in a Linux container with limited memory and storage resources.

• Container specifications: 8 virtual CPUs, RAM = 32 GB, the dataset will be mounted read-only into the container.

• Wall clock time for index construction: 24 hours.

• Minimum recall to be considered in the final ranking: 0.8.

• Search performance will be evaluated using 30 different search hyperparameters.

Task 3: Memory-Constrained Indexing without Reranking

In this task, participants are asked to develop memory-efficient indexing solutions that will be used without reranking the search results. The container provided will have higher memory capacity compared to Task 2. Participants have to build an index in the first phase. In the search phase, the original vectors cannot be used.

• Container specifications: RAM = 64 GB.

• Wall clock time for index construction: 12 hours.

• Minimum recall to be considered in the final ranking: 0.4

• Search performance will be evaluated using 30 different search hyperparameters.

Datasets and queries

• Input dataset: The 100M subset of the LAION2B dataset, i.e., see clip768 files.

• Public queries: The 10K private queries from the previous year's challenge will be used as the public query set. Queries asks for the 30nn ($k=30$).

• Private queries: An undisclosed set of 100K queries will be used for the final evaluation. Queries will be asking for the 30nn ($k=30$).

• Similarity measure: Dot product similarity.

• Output: two $k\times |queries|$ matrices in a single HDF5 filename (one file per search hyperparameter probe)

  • knns matrix to store object identifiers (integers)

  • dists matrix to store distances. Note that results should follow column-major order, i.e., the $i$th columns should contain identifiers and distances of the approximate $k$ nearest neighbors of the $i$th query sorted by the distance in ascending order, plus additional metadata. Also, knns identifiers must follow 1-based indexing (0-based indexing solutions must add 1 to evaluate correctly).

  • see examples for more details about these matrices and required metadata.