Residual Quantization with Implicit Neural Codebooks
This work addresses the need for more efficient and accurate vector quantization in data compression and search applications, representing an incremental improvement over existing methods.
The paper tackles the problem of residual quantization in vector search by proposing QINCo, a neural variant that constructs specialized codebooks per step based on previous approximations, resulting in improved accuracy over state-of-the-art methods, such as achieving better nearest-neighbor search with 12-byte codes than UNQ with 16 bytes on datasets like BigANN1M and Deep1M.
Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.