Spreading vectors for similarity search

Authors: Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Hervé Jégou

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section presents our experimental results. We focus on the class of similarity search methods that represents the database vectors with a compressed representation... All experiments have two phases.
Researcher Affiliation Collaboration Facebook AI Research Inria
Pseudocode No The paper describes algorithmic steps but does not present them in a structured pseudocode block or explicitly label any section as 'Algorithm'.
Open Source Code Yes The code is available online1. 1https://github.com/facebookresearch/spreadingvectors
Open Datasets Yes We use two benchmark datasets Deep1M and Big Ann1M. Deep1M consists of the first million vectors of the Deep1B dataset (Babenko & Lempitsky, 2016). We also experiment with the Big Ann1M (J egou et al., 2011b), which consists of SIFT descriptors (Lowe, 2004).
Dataset Splits Yes Both datasets contain 1M vectors that serve as a reference set, 10k query vectors and a very large training set of which we use 500k elements for training, and 1M vectors that we use a base to cross-validate the hyperparameters dout and λ.
Hardware Specification Yes All timings are for Big Ann1M are on a 2.2 GHz machine with 40 threads.
Software Dependencies No The paper mentions using 'Faiss (Johnson et al., 2017) implementation of PQ and OPQ' but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes Our model is a 3 layer perceptron, with Re LU non-linearity and hidden dimension 1024. The final linear layer projects the dataset to the desired output dimension dout, along with ℓ2-normalization. We use batch normalization (Ioffe & Szegedy, 2015) and train our model for 300 epochs with Stochastic Gradient Descent, with an initial learning rate of 0.1 and a momentum of 0.9. The learning rate is decayed to 0.05 (resp. 0.01) at the 80-th epoch (resp. 120-th).