This is a repo containing experiments concerning approximate nearest neighbors search techniques (ANN, ANNS, Similarity search)

Imagine you have to search a hudge music database to find songs most similar to song that you have just listened to. It would require too much time and resources to check every single song in the database for similarity and you are a bit impatient! That's where approximate search pops in. The idea is not to find exact nearest neighbour (although it might really find one), but to find those that are "close enough". Therefore the database is usually split into chunks, and real search happens in chunks that are likley to contain songs you want to hear next. Making and searching those chunks - that is what AAN techniques are about!

Theoretical ANN	Space partitioning example
The approximate search would add an ε error tolerance when finding nearest neighbours,but not so many real world implementations guarantee to find numbers in a given error range.	Imagine that algorithm divided space in 4 chunks. Searching for 3 closest neighbours would be pretty wrong if we searched blue space only. Increasing the number of partitions to search increases precision but requires more time.

This guy is gathering a nice protion of ANN techniques and making comparison between them:https://github.com/erikbern/ann-benchmarks These people have a cool library and implementations of techniques, checkout HNSW idea: https://github.com/nmslib/nmslib