Approximate Nearest Neighbor Search - Week 5

Last week, I have been working on the benchmarking system [1]. After considering different options, I created a new view to plot the progress of a specific metric for different values of a method parameter. For example, for knn, it is possible to analyze the number of base cases and runtime for different values of approximation error (epsilon), with different libraries/configurations.

I executed some tests for: mlpack_knn with cover trees and kdtrees, and other libraries like ANN and FLANN. Results are available on: [2] ("Metric analysis with multiple parameters for an algorithm/dataset combination.")

We plan to benchmark approximate neighbor search with bigger datasets, using Jenkins's servers.

Next week, I will continue working in the implementation of Spill Trees [3]. I will consider different approaches to include this extension and analyse the possibility of implementing Hybrid SP-Tree Search as a dual tree algorithm.