[Paper Review] Distill-VQ: Learning Retrieval Oriented Vector Quantization By Distilling Knowledge from Dense Embeddings

Shitao Xiao, Zheng Liu+11 more|arXiv (Cornell University)|Apr 1, 2022

Advanced Image and Video Retrieval Techniques4 citations

TL;DR

Distill-VQ proposes a knowledge distillation framework that trains vector quantization (VQ) modules—IVF and PQ—to preserve retrieval performance by mimicking relevance predictions from dense embeddings, achieving state-of-the-art recall and MRR on MS MARCO and Natural Questions without requiring labeled queries, even under high compression and low latency.

ABSTRACT

Vector quantization (VQ) based ANN indexes, such as Inverted File System (IVF) and Product Quantization (PQ), have been widely applied to embedding based document retrieval thanks to the competitive time and memory efficiency. Originally, VQ is learned to minimize the reconstruction loss, i.e., the distortions between the original dense embeddings and the reconstructed embeddings after quantization. Unfortunately, such an objective is inconsistent with the goal of selecting ground-truth documents for the input query, which may cause severe loss of retrieval quality. Recent works identify such a defect, and propose to minimize the retrieval loss through contrastive learning. However, these methods intensively rely on queries with ground-truth documents, whose performance is limited by the insufficiency of labeled data. In this paper, we propose Distill-VQ, which unifies the learning of IVF and PQ within a knowledge distillation framework. In Distill-VQ, the dense embeddings are leveraged as "teachers", which predict the query's relevance to the sampled documents. The VQ modules are treated as the "students", which are learned to reproduce the predicted relevance, such that the reconstructed embeddings may fully preserve the retrieval result of the dense embeddings. By doing so, Distill-VQ is able to derive substantial training signals from the massive unlabeled data, which significantly contributes to the retrieval quality. We perform comprehensive explorations for the optimal conduct of knowledge distillation, which may provide useful insights for the learning of VQ based ANN index. We also experimentally show that the labeled data is no longer a necessity for high-quality vector quantization, which indicates Distill-VQ's strong applicability in practice.

Motivation & Objective

• Address the misalignment between traditional vector quantization (VQ) objectives—minimizing reconstruction loss—and actual retrieval performance.
• Overcome the reliance of existing retrieval-optimized VQ methods on costly, limited labeled query-document pairs for contrastive learning.
• Enable effective training of VQ modules using massive unlabeled data by leveraging dense embeddings as knowledge teachers.
• Improve both IVF and PQ components jointly through distillation to enhance retrieval quality under high compression and low latency.
• Demonstrate that labeled data is no longer necessary for high-quality vector quantization in retrieval systems.

Proposed method

• Use pre-trained dense embeddings as 'teachers' to predict relevance scores between queries and candidate documents.
• Treat VQ modules (IVF and PQ) as 'students' that are trained to reproduce the teacher's relevance predictions.
• Apply ListNet loss to enforce ranking-order invariance between the teacher's and student's predicted relevance scores.
• Sample documents from Top-K and in-batch candidates to construct knowledge distillation signals without requiring ground-truth labels.
• Jointly optimize IVF and PQ components within a unified distillation framework to preserve retrieval performance.
• Integrate the trained VQ modules into FAISS's IVFPQ index for efficient inference and evaluation.

Experimental results

Research questions

• RQ1 Can knowledge distillation from dense embeddings effectively improve vector quantization performance without relying on labeled query-document pairs?
• RQ2 How does distilling relevance predictions from dense embeddings compare to minimizing reconstruction or retrieval loss directly in VQ training?
• RQ3 To what extent can Distill-VQ outperform existing VQ methods in terms of recall and MRR under high compression and low latency?
• RQ4 What is the impact of distillation on both IVF and PQ components, and how does it affect the efficiency-accuracy trade-off?
• RQ5 Can Distill-VQ achieve state-of-the-art performance even when labeled data is scarce or unavailable?

Key findings

• Distill-VQ significantly outperforms SOTA VQ methods on both MS MARCO and Natural Questions benchmarks in terms of Recall and MRR.
• At low bit-rates (e.g., 192× compression), Distill-VQ maintains substantial advantages in Recall@100 and MRR@10, especially under high compression and low latency.
• The method achieves better efficiency-recall trade-offs than FAISS’s original IVFPQ, particularly when codebook size is small.
• Even with small codebooks (e.g., 8×8), Distill-VQ outperforms baselines with larger codebooks, indicating strong generalization and data efficiency.
• The use of Top-K+In-Batch sampling enables effective distillation without labeled data, making the method scalable and practical.
• Distill-VQ maintains consistent performance gains across different compression ratios and search settings, demonstrating robustness and generalization.