Beyond the flat indexes that perform exhaustive searches, FAISS also has methods that compress the vectors to decrease their memory footprint. To accomplish this, FAISS has very efficient implementations of a few basic components like K-means, PCA, and **Product Quantizer encoding decoding. **We can use these components solely for the functions they provide, but they are usually used in conjunction with other methods.
We’ve already seen how PCA can be used in Part 1 and here we will look at indexing based on the Product Quantization (PQ) vector compression algorithm. These indexes do not use tree-based indexes, but they achieve the speeds in distance calculations by approximating and largely simplifying the distance calculations.
#machine-learning #ai #artificial-intelligence #data-science