Learning to Explain: An Information-Theoretic Perspective on Model Interpretation
Authors: Jianbo Chen, Le Song, Martin Wainwright, Michael Jordan
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We develop an efficient variational approximation to the mutual information, and show the effectiveness of our method on a variety of synthetic and real data sets using both quantitative metrics and human evaluation. |
| Researcher Affiliation | Collaboration | 1University of California, Berkeley 2Work done partially during an internship at Ant Financial 3Georgia Institute of Technology 4Ant Financial 5The Voleon Group. |
| Pseudocode | No | No structured pseudocode or algorithm blocks are provided in the paper. |
| Open Source Code | Yes | Codes for reproducing the key results are available online at https://github.com/Jianbo-Lab/ L2X. |
| Open Datasets | Yes | The Large Movie Review Dataset (IMDB) is a dataset of movie reviews for sentiment classification (Maas et al., 2011). The MNIST data set contains 28 28 images of handwritten digits (Le Cun et al., 1998). |
| Dataset Splits | No | The Large Movie Review Dataset (IMDB)... with a split of 25, 000 for training and 25, 000 for testing. The MNIST data set... with 11, 982 images for training and 1, 984 images for testing. No explicit validation split information (percentage or count) provided for these datasets. For synthetic data, it mentions 'median ranks of selected features for each sample in a validation data set are reported' but no explicit split details for this validation set. |
| Hardware Specification | Yes | all experiments were performed on a single NVidia Tesla k80 GPU |
| Software Dependencies | No | all experiments were performed on a single NVidia Tesla k80 GPU, coded in Tensor Flow. The version number for TensorFlow is not specified. |
| Experiment Setup | Yes | For all experiments, we use RMSprop (Maddison et al., 2016) with the default hyperparameters for optimization. We also fix the step size to be 0.001 across experiments. The temperature for Gumbel-softmax approximation is fixed to be 0.1. |