Improving Local Identifiability in Probabilistic Box Embeddings

Authors: Shib Dasgupta, Michael Boratko, Dongxu Zhang, Luke Vilnis, Xiang Li, Andrew McCallum

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply our model to a variety of synthetic problems and demonstrate that the new objective function correctly solves cases that are not solvable by any previous approaches to fitting box models. In real-world data, we demonstrate improved performance on a Word Net completion task and a Movie Lens density estimation task.
Researcher Affiliation Academia Shib Sankar Dasgupta Department of Computer Science University of Massachusetts, Amherst ssdasgupta@cs.umass.edu
Pseudocode Yes For ease of understanding, we provide a concrete instantiation of our algorithm for learning from pairwise conditional probabilities, incorporating all approximations, in Appendix D.
Open Source Code Yes Source code and data for the experiments are available at https://github.com/iesl/ gumbel-box-embeddings.
Open Datasets Yes Most of the recent geometric embeddings methods [26, 13, 5, 10, 18] use Word Net [14] to demonstrate the capability of their model to naturally represent real-world hierarchical data.
Dataset Splits No The paper discusses training and testing, but does not explicitly provide details about a separate validation split (e.g., percentages, sample counts, or cross-validation scheme) in the main text.
Hardware Specification No The paper does not specify the hardware used for running experiments (e.g., specific GPU or CPU models, memory details).
Software Dependencies No We use Weights & Biases package [2] to manage our experiments. The paper cites the software but does not specify its version number or other software dependencies with versions.
Experiment Setup No We train the embeddings via gradient descent, using KL-divergence loss. The paper mentions this general training approach and refers to appendices for 'full training details', but concrete hyperparameters are not detailed in the main text.