Self-explaining deep models with logic rule reasoning
Authors: Seungeon Lee, Xiting Wang, Sungwon Han, Xiaoyuan Yi, Xing Xie, Meeyoung Cha
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our method gives explanations closer to human decision logic than other methods while maintaining the performance of deep learning models. |
| Researcher Affiliation | Collaboration | Seungeon Lee KAIST School of Computing IBS Data Science Group archon159@kaist.ac.kr; Xiting Wang Social Computing Group Microsoft Research Asia xitwan@microsoft.com; Sungwon Han KAIST School of Computing IBS Data Science Group lion4151@kaist.ac.kr; Xiaoyuan Yi Social Computing Group Microsoft Research Asia xiaoyuanyi@microsoft.com; Xing Xie Social Computing Group Microsoft Research Asia xing.xie@microsoft.com; Meeyoung Cha IBS Data Science Group KAIST School of Computing mcha@ibs.re.kr |
| Pseudocode | No | The paper describes algorithms and formulations mathematically and in narrative text but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | Codes are released at Github.3 [https://github.com/archon159/SELOR] |
| Open Datasets | Yes | Yelp classifies reviews of local businesses into positive or negative sentiment [44], and Clickbait News Detection from Kaggle labels whether a news article is a clickbait [45]. Adult from the UCI machine learning repository [46], is an imbalanced tabular dataset that provides labels about whether the annual income of an adult is more than $50K/yr or not. |
| Dataset Splits | Yes | For Yelp, we use a down-sampled subset (10%) for training, as per existing work [39]. ... For Clickbait News Detection, we randomly selected 50% of the training data for the validation set and 50% for the test set. ... Adult: 32561 samples, 80% for training, 10% for validation, and 10% for testing. |
| Hardware Specification | Yes | All experiments were conducted on a single NVIDIA Quadro RTX 8000 GPU. |
| Software Dependencies | No | The paper mentions deep learning models and frameworks (e.g., BERT, RoBERTa) and an optimizer (Adam), but it does not specify software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1). |
| Experiment Setup | Yes | We use Adam optimizer [55] with a learning rate of 1e-4 and a batch size of 16 for all datasets. We train the model for 20 epochs for Yelp and Clickbait and 50 epochs for Adult. We use a warmup rate of 0.1 for the learning rate schedule. We use a weight decay of 0.01 for Yelp and Clickbait, and 0.001 for Adult. |