Interpretable Geometric Deep Learning via Learnable Randomness Injection
Authors: Siqi Miao, Yunan Luo, Mia Liu, Pan Li
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This work proposes a general mechanism, learnable randomness injection (LRI), which allows building inherently interpretable models based on general GDL backbones. LRI-induced models, once trained, can detect the points in the point cloud data that carry information indicative of the prediction label. We also propose four datasets from real scientific applications that cover the domains of high-energy physics and biochemistry to evaluate the LRI mechanism. Compared with previous post-hoc interpretation methods, the points detected by LRI align much better and stabler with the ground-truth patterns that have actual scientific meanings. LRI is grounded by the information bottleneck principle, and thus LRI-induced models are also more robust to distribution shifts between training and test scenarios. Our code and datasets are available at https://github.com/Graph-COM/LRI. |
| Researcher Affiliation | Academia | 1Georgia Institute of Technology, 2Purdue University |
| Pseudocode | No | The paper describes methods in text and uses diagrams, but does not contain formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and datasets are available at https://github.com/Graph-COM/LRI. |
| Open Datasets | Yes | We also propose four datasets from real scientific applications that cover the domains of high-energy physics and biochemistry to evaluate the LRI mechanism. ... These datasets cover important applications in HEP and biochemistry. ... Syn Mol. We utilize the molecules in ZINC (Irwin et al., 2012)... PLBind. We utilize protein-ligand complexes from PDBBind (Liu et al., 2017)... |
| Dataset Splits | Yes | Finally, we randomly split the dataset into training/validation/test sets with a ratio of 70 : 15 : 15. ... Finally, we randomly split the dataset into training/validation/test sets with a ratio of 70 : 15 : 15. ... Finally, we split the dataset into training/validation/test sets in a way that the number of molecules with or without either of these functional groups is uniformly distributed following (Mc Closkey et al., 2019) so that the dataset bias is minimized. ... Finally, we split the dataset into training/validation/test sets according to the year the complexes are discovered following St ark et al. (2022). |
| Hardware Specification | No | The paper does not specify the hardware (e.g., CPU, GPU models, memory) used for running the experiments. It only mentions general training parameters like batch size and optimizer settings. |
| Software Dependencies | No | All backbone models utilize implementations available in Pytorch-Geometric (Py G) (Fey & Lenssen, 2019). ... Both ETKDG and MMFF94 are implemented using RDKit. ... z µµ events are simulated with PYTHIA generator (Bierlich et al., 2022) overlaid with soft QCD pileup events, and particle tracks are simulated using Acts Common Tracking Software (Ai et al., 2022). The paper mentions software and frameworks used but does not provide specific version numbers for these dependencies (e.g., PyTorch, RDKit, Pythia, PyG versions). |
| Experiment Setup | Yes | We use a batch size of 128 on all datasets, except on Tau3Mu we use a batch size of 256 due to its large dataset size. The Adam (Kingma & Ba, 2015) optimizer with a learning rate of 1.0 10 3 and a weight decay of 1.0 10 5 are used. Acts Track, Syn Mol, and PLBind construct k-nn graphs with k being 5; Tau3Mu constructs the graph by drawing edges for any pair of points with a distance less than 1. For a fair comparison, all models will be trained with 300 epochs on Acts Track and Syn Mol and will be trained with 100 epochs on Tau3Mu and PLBind, so that all models are converged. ... For LRI-Bernoulli, α is tuned in {0.5, 0.7}. ... β is tuned in {1.0, 0.1, 0.01} for both methods after normalization based on the total number of points. |