Repository for the paper "von Mises-Fisher Sampling of GloVe Vectors" by W. Bendada, G. Salha-Galvan, R. Hennequin, T. Bontempelli, T. Bouabça and T. Cazenave, presented at the FPI workshop of ICLR 2025.
We recently introduced, in [1], von Mises-Fisher exploration (vMF-exp), a scalable sampling method for exploring large action sets in reinforcement learning problems where hyperspherical embedding vectors represent these actions. In this prior work, we demonstrated that vMF-exp scales to millions of actions and exhibits several desirable properties. However, we did not test vMF-exp on publicly available real-world data, which is essential for reproducibility and deeper understanding of the method.
We address this limitation by experimentally validating the main properties of vMF-exp on a large-scale, publicly available real-world dataset of GloVe word embedding vectors. The purpose of this paper is to provide a fresh perspective on our initial work, with reinforced validation.
The GloVe word embedding dataset used in our experiments is publicly available for download at: https://nlp.stanford.edu/projects/glove/
Experiences were run using the 25-dimensional embedding vectors (GloVe-25).
After downloading the correct file, unzip it and place it in a folder named dataset.
The script compute_probas.py will run Monte Carlo simulations estimating the probability for von Mises-Fisher exploration and Boltzmann exploration to sample an action with known similarity given a state vector.
All vectors are sampled from the GloVe-25 dataset previously downloaded. The result can then be plotted using plot_probas.py.
For instance, to reproduce Figure 2.a, one can run the following command:
python -m src.compute_probas -k 1 -a 0.0 -n glove.25 -bs 3000 -nt 10000
which will run the corresponding Monte Carlo Simulations, followed by the command:
python -m src.plot_probas --path results/glove.25/k\=1.0_a\=0.00_samples\=30000000/
which will create a plot similar to the following one:
and saved in a sub-folder of /results/ named according to the chosen parameters.
The script compare_boltzmann_vs_vmf.py will reproduce Figure 1.a and Figure 1.b for a specified range of values of <V,A> that must first be computed using compute_probas.py with changing values of a (see above).
For instance, running compute_probas.py several times with values of a in [0.9,0.3,0.0,-0.3,-0.9] and then running:
python -m src.compare_boltzmann_vs_vmf --values 0.9,0.3,0.0,-0.3,-0.9
will create the following two plots:
[1] Bendada et al., vMF-exp: von Mises-Fisher Exploration of Large Action Sets with Hyperspherical Embeddings, 2024