Incorporating Inductive Biases to Energy-based Generative Models
Authors: Yukun Li, Liping Liu
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the effectiveness of this new approach on three tasks: modeling molecular data, fitting handwritten digits, and modeling point clouds. The results demonstrate the effectiveness of the proposed approach in a wide range of data-fitting domains. 5 Experiments We evaluate the effect of our special statistics on data fitting through three generative tasks: molecule graph generation, image generation, and point cloud generation. Experiments related to molecule generation are detailed in Section 5.1. Section 5.2 discussed the image generation experiments. Section 5.4 discussed the point cloud generation task. |
| Researcher Affiliation | Academia | Yukun Li EMAIL Department of Computer Science Tufts University Li-Ping Liu EMAIL Department of Computer Science Tufts University |
| Pseudocode | No | The paper describes the hybrid energy-based model and its training procedure using score matching but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. It mentions using existing libraries like RDKit but makes no statement about releasing its own implementation. |
| Open Datasets | Yes | Our model is evaluated on the QM9 dataset (Pinheiro et al., 2020), which comprises a diverse collection of 133,885 molecules... We evaluated the effectiveness of our method on the MNIST dataset, a widely used benchmark for various image downstream tasks... We tested our approach on the Fashion MNIST dataset Kayed et al. (2020)... We evaluated our model on the Shape Net dataset (Chang et al., 2015), a widely used benchmark in 3D shape analysis and understanding. |
| Dataset Splits | No | The paper mentions evaluating models on 'test set' for various tasks (e.g., in Section 5.2 and 5.3), implying data splitting. However, it does not provide specific details on the split percentages, sample counts, or the methodology used to create these splits. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | Following previous studies, the molecules are kekulizated using the RDKit library (Bento et al., 2020), and hydrogen atoms are removed. |
| Experiment Setup | Yes | We trained GDSS for 300 epochs, the batch size is 1024, and the learning rate is 5e-3. We trained EDP-GNN for 1000 epochs, the batch size is 1024, and the learning rate is 8e-3. (Section 5.1). The model is trained for 1000 epochs... The batch size is 4096, and the learning rate is 1e-2. The learning rate is kept constant for the first 300 epochs and decreases linearly from 300 to 1000 epochs. (Section 5.2). We evaluated performance using the VE-SDE method with a batch size of 64 and an initial learning rate 1e-2. The learning rate decays step-wise by multiplying it by 0.95 every ten epochs. The model was trained for 100 epochs. (Section 5.3). |