Learning of Discrete Graphical Models with Neural Networks
Authors: Abhijith Jayakumar, Andrey Lokhov, Sidhant Misra, Marc Vuffray
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3.2 Experiments. Now we will test Neur ISE on two highly structured graphical models. In our testing we will compare Neur ISE to GRISE. Figure 1: Learning the model given by Eq. (16), with p = 10 and L = 6. (a) ℓ1 error in the learned conditionals averaged over all possible inputs. (b) The absolute value of the leading coefficient of the learned model at each order compared to that of the true model. |
| Researcher Affiliation | Collaboration | Abhijith J. Centre for High Energy Physics, Indian Institute of Science, Bengaluru. abhijithj@iisc.ac.in. Andrey Y. Lokhov, Sidhant Misra, Marc Vuffray Theoretical Division, Los Alamos National Laboratory { lokhov, sidhant, vuffray }@lanl.gov |
| Pseudocode | No | The paper describes the Neur ISE algorithm and its optimization but does not contain structured pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | The code and data for this can be found at https://github.com/lanl-ansi/Neur ISE. |
| Open Datasets | No | The paper states it 'generate[s] samples from a graphical model' for its experiments. While the code and data are available at a GitHub link, this refers to data they generated themselves, not a distinct, publicly available, or open dataset in the sense of a named benchmark or a dataset with its own dedicated access information beyond their experiment's repository. |
| Dataset Splits | No | The paper refers to 'training samples' and 'testing samples' but does not specify any dataset splits, percentages, or absolute counts for training, validation, or test sets, nor does it mention a cross-validation setup. |
| Hardware Specification | No | The paper mentions that 'the learning process in this case can be easily parallelized on a GPU using off-the-shelf machine learning libraries' and that 'on real hardware the performance of these will depend on implementation'. However, it does not provide specific hardware details such as GPU models (e.g., NVIDIA A100), CPU models, or memory specifications. |
| Software Dependencies | No | We will be using feed forward neural nets with the swish activation function (swish(x) = x sigmoid(x)) [25] . We specify the size of a neural net with two numbers, d and w, which will be the number of hidden layers in the model and the number of neurons in each hidden layer respectively. All the nets were trained using the ADAM optimizer [19]. However, no specific version numbers are provided for these software components. |
| Experiment Setup | Yes | We specify the size of a neural net with two numbers, d and w, which will be the number of hidden layers in the model and the number of neurons in each hidden layer respectively. All the nets were trained using the ADAM optimizer [19]. At the beginning of training the input weights must be initialized to zero. Secondly, we must regularize the Neur ISE loss function with the ℓ1 norm of only the input weights. |