Binary Hyperbolic Embeddings
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Hyperbolic, Binary, Hierarchical
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We enable fast retrieval in hyperbolic space through binarization at low bit rates.
Abstract: As datasets continue to grow, vector-based search becomes more storage and compute intensive, requiring large-scale systems to support retrieval. Proposed solutions range from quantization techniques that balance speed and accuracy, to hashing methods that learn compact binary representations. This paper promotes the use of hyperbolic space for its compact nature whilst overcoming its slow retrieval via binarization. Specifically, we address hyperbolic space's inherent slowness by proving that its complex similarity calculations can be equated to a binary XOR operation. Our approach allows for 90% less storage and at least 4.7 times faster search while maintaining performance of full-precision Euclidean embeddings.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1849
Loading