Tractable Regularization of Probabilistic Circuits

Authors: Anji Liu, Guy Van den Broeck

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that both methods consistently improve the generalization performance of a wide variety of PCs. Moreover, when paired with a simple PC structure, we achieved state-of-the-art results on 10 out of 20 standard discrete density estimation benchmarks. Open-source code and experiments are available at https://github.com/UCLA-Star AI/Tractable-PC-Regularization.
Researcher Affiliation Academia Anji Liu Department of Computer Science UCLA Los Angeles, CA 90095 liuanji@cs.ucla.edu Guy Van den Broeck Department of Computer Science UCLA Los Angeles, CA 90095 guyvdb@cs.ucla.edu
Pseudocode Yes Algorithm 1 Forward pass ... Algorithm 2 Backward pass ... Algorithm 3 PC Entropy regularization
Open Source Code Yes Open-source code and experiments are available at https://github.com/UCLA-Star AI/Tractable-PC-Regularization.
Open Datasets Yes Empirical evaluation We empirically evaluate both proposed regularization methods on the twenty density estimation datasets [39]... We first examine the performance on a protein sequence dataset [29] that suffers from severe overfitting.
Dataset Splits No While the paper mentions the use of 'validation set' multiple times (e.g., 'using the validation set and report results on the test set'), it does not provide specific details on how these splits were performed (e.g., percentages or sample counts for training, validation, and test sets in the main text).
Hardware Specification No The main text of the paper does not specify any hardware details. While the checklist mentions 'Details about computing resources can be found in Appendix B.3', this appendix is not provided in the given text.
Software Dependencies No The paper does not provide specific software names with version numbers in the main text. While the checklist mentions 'All details for reproducibility are specified in Appendices B.3 and B.4', these appendices are not provided in the given text.
Experiment Setup Yes For all experiments, we performed a hyperparameter search for all three regularization approaches (Laplace smoothing, data softening, and entropy regularization)5 using the validation set and report results on the test set. ... 5Specifically, α {0.1, 0.4, 1.0, 2.0, 4.0, 10.0}, β {0.9996, 0.999, 0.996}, τ {0.001, 0.01, 0.1}. ... For all experiments, we trained the PCs with 100 mini-batch EM epochs and 100 full-batch EM epochs.