Abstract:Vector quantization techniques, such as Product Quantization (PQ), play a vital role in approximate nearest neighbor search (ANNs) and maximum inner product search (MIPS) owing to their remarkable search and storage efficiency. However, the indexes in vector quantization cannot be trained together with the inference models since data indexing is not differentiable. To this end, differentiable vector quantization approaches, such as DiffPQ and DeepPQ, have been recently proposed, but existing methods have two drawbacks. First, they do not impose any constraints on codebooks, such that the resultant codebooks lack diversity, leading to limited retrieval performance. Second, since data indexing resorts to operator, differentiability is usually achieved by either relaxation or Straight-Through Estimation (STE), which leads to biased gradient and slow convergence. To address these problems, we propose a Differentiable Optimized Product Quantization method (DOPQ) and beyond in this paper. Particularly, each data is projected into multiple orthogonal spaces, to generate multiple views of data. Thus, each codebook is learned with one view of data, guaranteeing the diversity of codebooks. Moreover, instead of simple differentiable relaxation, DOPQ optimizes the loss based on direct loss minimization, significantly reducing the gradient bias problem. Finally, DOPQ is evaluated with seven datasets of both recommendation and image search tasks. Extensive experimental results show that DOPQ outperforms state-of-the-art baselines by a large margin.

Codebook-softened product quantization for high accuracy approximate nearest neighbor search

Product quantization with dual codebooks for approximate nearest neighbor search

Flexible Product Quantization for Fast Approximate Nearest Neighbor Search.

PQBF: I/O-Efficient Approximate Nearest Neighbor Search by Product Quantization.

Derived Codebooks for High-Accuracy Nearest Neighbor Search

Mean-Removed Product Quantization for Approximate Nearest Neighbor Search.

Stacked Product Quantization For Nearest Neighbor Search On Large Datasets

Dynamic programming based optimized product quantization for approximate nearest neighbor search.

A New Cell-Level Search Based Non-Exhaustive Approximate Nearest Neighbor (ANN) Search Algorithm in the Framework of Product Quantization.

Quarter-Point Codeword Expansion for Product Quantization.

Quarter-Point Product Quantization for approximate nearest neighbor search

Semi-supervised product quantization for approximate nearest neighbor search

Transformed Residual Quantization for Approximate Nearest Neighbor Search

Deep Product Quantization Module for Efficient Image Retrieval

Optimized High Order Product Quantization for Approximate Nearest Neighbors Search.

Local Deep Learning Quantization for Approximate Nearest Neighbor Search

Distribution Sensitive Product Quantization.

Information Gain Product Quantization for Image Retrieval

Scalable Image Retrieval by Sparse Product Quantization

Multi-Pqtable For Approximate Nearest-Neighbor Search

Differentiable Optimized Product Quantization and Beyond