BATL: Learning Balanced Tree Indexes for Large-Scale Vector Retrieval скачать в хорошем качестве

BATL: Learning Balanced Tree Indexes for Large-Scale Vector Retrieval

Paper: https://doi.org/10.1145/3580305.3599406 In a nutshell: Model Definition: Encoder-decoder Transformer predicting branch-ID sequences (paths) on a balanced K-ary tree to locate nearest-neighbor buckets. Reduced Complexity: Flattens massive multi-class bucket classification into small-vocabulary autoregressive routing, improving generalization and training efficiency. Training Loop: Initialize random balanced tree - train seq2seq model - update tree structure by re-assigning data points layer-by-layer based on routing likelihood - repeat until convergence. Training Signal: Query paired with the branch-ID sequence of its ground-truth nearest neighbors; neighbors identified via exhaustive search or subset sampling on large-scale databases. Failure Modes & Fixes: Lop-sided partitions - enforce branch capacity limit αN/K^h. Collapse (vectors in few buckets) - check routing decision in descending order and assign to first branch satisfying balance constraints. Error accumulation in deep trees - use low H (height) to minimize layer-wise prediction drift. Inference-Time Safeguards: Beam search navigates the encoder-decoder to identify the top-M most likely candidate paths/buckets. Boundary / Coverage Issues: Data near bucket boundary missed - ensemble multiple diversified trees and independent models to merge candidate set results. Candidate Selection / Ranking: Merge points from top-M buckets - filter out points appearing only once to reduce scale - compute exact similarity - final ranking for top-k. Key Parameters: α (load flexibility/balance factor), K (branching factor/vocabulary size), H (tree height), beam size (search breadth). Efficiency Considerations: Sublinear query time; O(TN) memory usage; shared branch embeddings across layers significantly reduce model footprint versus flat MLP baselines. Intuition: Breaking one massive "which bucket?" guess into a sequence of small "which branch?" decisions makes the index easier for neural nets to learn and much faster to navigate. *Disclaimer*: This is an AI-powered production. The scripts, insights, and voices featured in this podcast are generated entirely by Artificial Intelligence models. While we strive for technical accuracy by grounding our episodes in original research papers, listeners are encouraged to consult the primary sources for critical applications.