Online Optimized Product Quantization for Dynamic Database Using SVD-Updating

Approximate nearest neighbor (ANN) search allows us to perform similarity search over massive vectors with less memory and computation. Optimized Product Quantization (OPQ) is one of the state-of-the-art methods for ANN where data vectors are represented as combinations of codewords by taking into account the data distribution. However, it suffers from degradation in accuracy when the database is frequently updated with incoming data whose distribution is different. An existing work, Online OPQ, addressed this problem, but the computational cost is high because it requires to perform of costly singular value decompositions for updating the codewords. To this problem, we propose a method for updating the rotation matrix using SVD-Updating, which can dynamically update the singular matrix using low-rank approximation. Using SVD-Updating, instead of performing multiple singular value decompositions on a high-rank matrix, we can update the rotation matrix by performing only one singular value decomposition on a low-rank matrix. In the experiments, we prove that the proposed method shows a better trade-off between update time and retrieval accuracy than the comparative methods.