Spectral Approaches to Nearest Neighbor Search
summary
Proposes reducing approximate nearest neighbor to a spectral graph problem, building a sparsified neighborhood graph and embedding via Laplacian eigenmaps so most k‑nearest relations remain (≈95%) after compression by about 10×. Implements a data structure that combines spectral sparsifiers, Laplacian eigenmaps and LSH, and evaluates on image and text collections, reporting roughly 5–10× query speedup with about 50% memory reduction compared to baseline ANN.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:45Nearest Neighbor Search (NNS)
- 1:08Motivation
- 3:18Curse of dimensionality
- 4:14Approximate NNS
- 16:37Practice vs Theory Data-aware projections often outperform (vanilla) random-projection methods
- 19:43Model properties
- 24:32Tool: PCA Principal Component Analysis like SVD Top PCA direction direction maximizing variance of the data
- 25:01Naive attempt via PCA Use PCA to find the signal subspace U?
- 27:291st Algorithm: intuition
- 30:51Analysis of PCA
- 32:51Closeness of subspaces ?
- 38:19Additional issue: Conditioning After an iteration, the noise is not random anymore!
- 40:06Performance of Iterative PCA
- 42:342 algorithmic modifications
- 44:35Post-perspective
related talks
43mSpectral Approaches to Nearest Neighbor Search
01h 05mNearest Neighbor Methods I
34mParallel Efficient Clustering with Graph-Based Approximate Nearest Neighbor Search
42mDistribution-Specific Analysis of Nearest Neighbor Search and Classification
48mHolder Homeomorphisms and Approximate Nearest Neighbors
46mSpectral Partitioning for Metrics (And NNs Too)
10man introduction to approximate nearest neighbors
40mEfficient Reductions for k-Nearest Neighbor Search
