The 2023 Big ANN Challenge, held at NeurIPS 2023, focused on advancing the state-of-the-art in indexing data structures and search algorithms for practical variants of Approximate Nearest Neighbor (ANN) search that reflect its the growing complexity and diversity of workloads. Unlike prior challenges that emphasized scaling up classical ANN search (Simhadri et al., NeurIPS 2021), this competition addressed sparse, filtered, out-of-distribution, and streaming variants of ANNS. Participants developed and submitted innovative solutions that were evaluated on new standard datasets with constrained computational resources. The results showcased significant improvements in search accuracy and efficiency, with notable contributions from both academic and industrial teams. This paper summarizes the competition tracks, datasets, evaluation metrics, and the innovative approaches of the top-performing submissions, providing insights into the current advancements and future directions in the field of approximate nearest neighbor search.

We develop a linearizable lock-free data structure, the CLEANN-tree (Concurrent Linearizable Efficient Augmented Nearest Neighbor tree), for low dimensional κ-nearest-neighbor searching. The data structure maintains a set of points P in d dimensions under insertion and deletion, while supporting queries that, given a point, return the k nearest points in P. The CLEANN-tree is constructed by modifying a kd-tree, a type of spatial decomposition commonly used for κ-nearest neighbor searching, for the concurrent environment. It is the first such concurrent structure---two previous structures were either not linearizable or only supported κ=1. Furthermore CLEANN-tree stores an augmented value (more specifically, a bounding box) in each internal node of the kd-tree. These bounding boxes significantly improve query performance by allowing more aggresive pruning. However, correctly and efficiently maintaining these augmented values is challenging in the linearizable lock-free setting because queries can examine large parts of the structure, which might be changing, and an insert or delete can require updating all the augmented values from the leaf to the root.

Vector indexing enables semantic search over diverse corpora and has become an important interface to databases for both users and AI agents. Efficient vector search requires deep optimizations in database systems. This has motivated a new class of specialized vector databases that optimize for vector search quality and cost. Instead, we argue that a scalable, high-performance, and cost-efficient vector search system can be built inside a cloud-native operational database like Azure Cosmos DB while leveraging the benefits of a distributed database such as high availability, durability, and scale. We do this by deeply integrating DiskANN, a state-of-the-art vector indexing library, inside Azure Cosmos DB NoSQL. This system uses a single vector index per partition stored in existing index trees, and kept in sync with underlying data. It supports < 20ms query latency over an index spanning 10 million vectors, has stable recall over updates, and offers approximately 43× and 12× lower query cost compared to Pinecone and Zilliz serverless enterprise products. It also scales out to billions of vectors via automatic partitioning. This convergent design presents a point in favor of integrating vector indices into operational databases in the context of recent debates on specialized vector databases, and offers a template for vector indexing in other databases.

Approximate nearest-neighbor search (ANNS) algorithms are a key part of the modern deep learning stack due to enabling efficient similarity search over high-dimensional vector space representations (i.e., embeddings) of data. Among various ANNS algorithms, graph-based algorithms are known to achieve the best throughput-recall tradeoffs. Despite the large scale of modern ANNS datasets, existing parallel graph-based implementations suffer from significant challenges to scale to large datasets due to heavy use of locks and other sequential bottlenecks, which 1) prevents them from efficiently scaling to a large number of processors, and 2) results in non-determinism that is undesirable in certain applications.

We study the connections between sorting and the binary search tree (BST) model, with an aim towards showing that the fields are connected more deeply than is currently appreciated. While any BST can be used to sort by inserting the keys one-by-one, this is a very limited relationship and importantly says nothing about parallel sorting. We show what we believe to be the first formal relationship between the BST model and sorting. Namely, we show that a large class of sorting algorithms, which includes mergesort, quicksort, insertion sort, and almost every instance-optimal sorting algorithm, are equivalent in cost to offline BST algorithms. Our main theoretical tool is the geometric interpretation of the BST model introduced by Demaine et al., which finds an equivalence between searches on a BST and point sets in the plane satisfying a certain property. To give an example of the utility of our approach, we introduce the log-interleave bound, a measure of the information-theoretic complexity of a permutation $π$, which is within a $\lg \lg n$ multiplicative factor of a known lower bound in the BST model; we also devise a parallel sorting algorithm with polylogarithmic span that sorts a permutation $π$ using comparisons proportional to its log-interleave bound. Our aforementioned result on sorting and offline BST algorithms can be used to show existence of an offline BST algorithm whose cost is within a constant factor of the log-interleave bound of any permutation $π$.

We present a set of parallel algorithms for computing exact k-nearest neighbors in low dimensions. Many k-nearest neighbor algorithms use either a kd-tree or the Morton ordering of the point set; our algorithms combine these approaches using a data structure we call the zd-tree. We show that this combination is both theoretically efficient under common assumptions, and fast in practice. For point sets of size n with bounded expansion constant and bounded ratio, the zd-tree can be built in O(n) work with O(nε) span for constant ε < 1, and searching for the k-nearest neighbors of a point takes expected O(k log k) time. We benchmark our k-nearest neighbor algorithms against existing parallel k-nearest neighbor algorithms, showing that our implementations are generally faster than the state of the art as well as achieving 75x speedup on 144 hyperthreads. Furthermore, the zd-tree supports parallel batch-dynamic insertions and deletions; to our knowledge, it is the first k-nearest neighbor data structure to support such updates. On point sets with bounded expansion constant and bounded ratio, a batch-dynamic update of size k requires O(k log n/k) work with O(kε + polylog(n)) span.

The Problem-Based Benchmark Suite (PBBS) is a set of benchmark problems designed for comparing algorithms, implementations and platforms. For each problem, the suite defines the problem in terms of the input-output relationship, and supplies a set of input instances along with input generators, a default implementation, code for checking correctness or accuracy, and a timing harness. The suite makes it possible to compare different algorithms, platforms (e.g. GPU vs CPU), and implementations using different programming languages or libraries. The purpose is to better understand how well a wide variety of problems parallelize, and what techniques/algorithms are most effective.

We consider settings where agents are evaluated based on observed features, and assume they seek to achieve feature values that bring about good evaluations. Our goal is to craft evaluation mechanisms that incentivize the agents to invest effort in desirable actions; a notable application is the design of course grading schemes. Previous work has studied this problem in the case of a single agent. By contrast, we investigate the general, multi-agent model, and provide a complete characterization of its computational complexity.