HORSE: Hierarchical Representation for Large-Scale Neural Subset Selection
Authors: Binghui Xie, Yixuan Wang, Yongqiang Chen, Kaiwen Zhou, Yu Li, Wei Meng, James Cheng
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experimentation, we demonstrate that HORSE significantly enhances neural subset selection performance by capturing more complex information and surpasses the state-of-the-art methods in handling large-scale inputs by a margin of up to 20%. |
| Researcher Affiliation | Academia | Binghui Xie, Yixuan Wang, Yongqiang Chen, Kaiwen Zhou, Yu Li, Wei Meng, James Cheng Department of Computer Science and Engineering The Chinese University of Hong Kong |
| Pseudocode | Yes | C Pseudo Code of HORSE In the main text, we present HORSE using matrix calculations, which may be challenging to comprehend. To improve understanding of how our method establishes an attention map between subsets Si, we detail the procedural steps in Algorithm 1, with a special emphasis on the generation of h(Si). |
| Open Source Code | Yes | We provided part of our code at the time of submission. Furthermore, our method is based on the open-source repository detailed in Ou et al. [2022]. We have also comprehensively documented our experimental settings in the Experiment section and the Appendix. The integration of the open-source repository with our code ensures the reproducibility of our empirical results. |
| Open Datasets | Yes | For this experiment, we utilize the dataset from the Amazon baby registry, sourced from Gillenwater et al. [2014a]... Our experiments are conducted on two datasets: PDBBind [Liu et al., 2015a] and Binding DB [Liu et al., 2007]. |
| Dataset Splits | Yes | We follow the procedure of Ou et al. [2022] to obtain 1,000 samples, subsequently divided into training, validation, and test sets. ... The remaining subsets are then divided into training, validation, and test folds using a 1:1:1 ratio. ... We constructed separate training, validation, and test splits, comprising 1000, 100, and 100 data points, respectively. |
| Hardware Specification | Yes | Notably, the memory capacity of the Ge Force RTX 3090 is insufficient when the size reaches 600. ... allows efficient training on a single Ge Force RTX 3090 GPU. |
| Software Dependencies | No | The proposed models are trained using the Adam optimizer [Kingma and Ba, 2014]... No specific version numbers for software dependencies or libraries are provided beyond the Adam optimizer citation. |
| Experiment Setup | Yes | The proposed models are trained using the Adam optimizer [Kingma and Ba, 2014] with a fixed learning rate of 1e 4 and a weight decay rate of 1e 5. To accommodate different model sizes across various datasets, we select the batch size from the set {4, 8, 16, 32, 64, 128}. ... if there is no performance improvement over 10 consecutive epochs, we terminate the training process prematurely. For each dataset, the maximum number of epochs allowed for training is set to 80. |