Paired-Dual Learning for Fast Training of Latent Variable Hinge-Loss MRFs

Authors: Stephen Bach, Bert Huang, Jordan Boyd-Graber, Lise Getoor

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate on social-group detection, trust prediction in social networks, and image reconstruction, finding that paired-dual learning trains models as accurate as those trained by traditional methods in much less time, often before traditional methods make even a single parameter update.
Researcher Affiliation Academia Stephen H. Bach University of Maryland, College Park, MD Bert Huang Virginia Tech, Blacksburg, VA Jordan Boyd-Graber University of Colorado, Boulder, CO Lise Getoor University of California, Santa Cruz, CA
Pseudocode Yes Algorithm 1 Paired-Dual Learning
Open Source Code No The paper does not contain any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We use the data of Bach et al. (2013a)...We evaluate on a subsample of roughly 2,000 users of Epinions.com (Huang et al., 2013; Richardson et al., 2003)...Using the 400-image Olivetti face data set...
Dataset Splits Yes measuring the area under the precision recall curve (Au PR) using ten folds of cross-validation...perform eight-fold cross-validation
Hardware Specification No The paper mentions 'avoiding confounding factors such as differences in hardware used in these experiments' but does not provide specific details on the hardware (e.g., GPU/CPU models, memory) used for the experiments.
Software Dependencies No The paper does not specify version numbers for any software dependencies or libraries used in the experiments (e.g., Python, PyTorch, specific solvers).
Experiment Setup Yes We test two variants of paired-dual learning: the finest grained interleaving with only two ADMM iterations per weight update (N = 1) and a coarser grained 20 ADMM iterations per update (N = 10)...In our experiments (Section 4), K = 0 often suffices, but for one task, using K = 10 produces a better start to optimization.