Reproduce “Revisiting Neural Retrieval on Accelerators” on OpenResearch

About this paper

Jiaqi Zhai, Zhaojie Gong, Yueming Wang, Xiao Sun, Zheng Yan, Fu Li, Xing Liu

This research introduces a hierarchical retrieval strategy that combines a novel high-rank Mixture of Logits (MoL) similarity function with accelerator-optimized infrastructure to enable expressive neural retrieval at scale. This approach yields substantial improvements in recommendation quality and reduced popularity bias on both public and industrial datasets, maintaining competitive online serving latency.

Key insights

• Expressive, high-rank similarity functions can be effectively deployed in large-scale retrieval systems without prohibitive latency by combining a hierarchical architecture with hardware-aware and distributed training optimizations.
• Mixture models, such as MoL, can capture significantly more complex user-item interactions (demonstrated by higher effective rank) than traditional dot-product or simple MLP models, leading to improved recommendation quality and better generalization.
• Specific regularization techniques, like component-level hypersphere embeddings and gating dropout, are crucial for ensuring the generalization and mitigating popularity bias in high-rank neural retrieval models.

Summary by alphaXiv.