#approximate-nearest-neighbors

Question 1: Single choice

mediumAI-checkedNearest Neighbors & Kernel Methods

Finding the 3 words most similar to "chip" by cosine similarity among 400,000 GloVe vectors compares the query with every vector. How can such queries be made much faster for a very large vocabulary?

Build an approximate nearest-neighbor index, e.g. locality-sensitive hashing or an HNSW graph1
Use Euclidean distance instead of cosine similarity, since it avoids normalization2
Ask for 1 neighbor instead of 3, since the scan cost grows with the number returned3
Query with $\mathrm{vec}(c) + \mathrm{vec}(b) - \mathrm{vec}(a)$ instead of the word's own vector4

#approximate-nearest-neighbors

Timed test