General Purpose MRF Learning with Neural Network Potentials

Authors: Hao Xiong, Nicholas Ruozzi

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate experimentally that our approach is capable of effectively modeling the data distributions of a variety of real data sets and that it can compete effectively with other common methods on multilabel classification and generative modeling tasks. 4 Experimental Results In this section we will illustrate the practical performance of our method.
Researcher Affiliation Academia Hao Xiong and Nicholas Ruozzi The University of Texas at Dallas {hao.xiong, nicholas.ruozzi}@utdallas.edu
Pseudocode Yes Algorithm 1: MLE with NN potentials
Open Source Code Yes All code and data used as part of these experiments will be made available on Git Hub.1 (Footnote 1: https://github.com/motionlife/nnmrf)
Open Datasets Yes on eight data sets from the UCI repository and compare the results with a variety of standard classifiers: Support Vector Machines (SVMs), Gaussian Naive Bayes (GNB), logistic regression (LR) , decision trees (DTs), random forrests (RFs), k-nearest neighbor (KNN), and multi-layer perceptrons (MLP). ... as well as the MNIST image classification data set [Le Cun et al., 1998]
Dataset Splits No Table 1 shows the average test accuracy of all methods over 30 randomly chosen 80/20, train/test splits. The paper specifies train/test splits but does not provide details for a validation split.
Hardware Specification No Our algorithm was implemented in Python using the Tensor Flow 2 framework to take advantage of fast GPU training and inference. As a result, the entire training process can be implemented quite efficiently on modern GPU accelerators. The paper mentions GPUs generally but does not specify exact models or configurations.
Software Dependencies No Our algorithm was implemented in Python using the Tensor Flow 2 framework to take advantage of fast GPU training and inference. For the competing methods, we used the implementations available in scikit-learn. The paper mentions software names but does not provide version numbers for them.
Experiment Setup Yes Both node and edge potentials are neural nets with 4 hidden layers and the hidden nodes for each layer are [ 3, 5, 5, 3 ] for each node potential, [ 5, 7, 7, 5 ] for each edge potential respectively, with all hidden layer s activation function set to be Se LUs (scaled exponential linear units) [Klambauer et al., 2017] and the output layer activation is Tan H. We used a belief pool size of T = 20, and we ran 100 iterations in each marginal inference procedure during learning with the ADAM optimizer.