XAI for Transformers: Better Explanations through Conservative Propagation
Authors: Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, Lior Wolf
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide qualitative and quantitative experiments on different data domains, such as natural language understanding, computer vision and graph analysis. Comparison is performed with state-of-the-art baseline explanation methods for Transformers. For the quantitative analysis, we perform different input perturbation schemes, in which we track the behavior of the model when relevant or irrelevant features are added to or removed from the input data sample. The proposed explanation method for Transformers presents excellent performance on the qualitative and quantitative experiments and outperforms all existing baseline methods in most of the tasks. |
| Researcher Affiliation | Academia | 1The School of Computer Science, Tel-Aviv University, Israel 2Machine Learning Group, Technische Universit at Berlin, Berlin, Germany 3BIFOLD Berlin Institute for the Foundations of Learning and Data, Berlin, Germany 4Department of Artificial Intelligence, Korea University, Seoul, Korea 5Max Planck Institute for Informatics, Saarbr ucken, Germany. |
| Pseudocode | No | The paper describes an "Implementation Trick" with step-by-step instructions, but it is not formally presented as pseudocode or an algorithm block with a clear label. |
| Open Source Code | Yes | Our code is publicly available.1 1https://github.com/Ameen Ali/XAI_ Transformers |
| Open Datasets | Yes | We use the following datasets from natural language processing, image classification, as well as molecular modeling to evaluate the different XAI approaches. For the NLP experiments, we consider sentiment classification on the SST-2 (Socher et al., 2013) and IMDB datasets (Maas et al., 2011)... For experiments with graph Transformers, the MNIST superpixels data (Monti et al., 2017)... Another graph dataset is based on the Molecule Net (Wu et al., 2018) benchmark: the BACE dataset... |
| Dataset Splits | No | The paper mentions "early stopping for decreasing validation performance" in Appendix C, implying the use of a validation set. However, it does not provide specific details on how this validation set was split (e.g., percentages or sample counts) to reproduce the partitioning. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, or memory) used to run its experiments. |
| Software Dependencies | No | The paper mentions using "pretrained BERT-Transformers" and the "Graphormer model" but does not provide specific version numbers for any underlying software dependencies (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | For training, we use batchsizes of bs = 32 and optimize the model parameters using the Adam W optimizer with a learning rate of lr = 2e 5 for a maximal number of T = 20 epochs or until early stopping for decreasing validation performance is reached. ... We train a 2-layer graphormer model with a batchsize of bs = 64 and Adam W as an optimizer with lr = 2e 4. The model is trained for 1000 epochs for the BACE dataset and 10 epochs for MNISTSuperpixels... |