Deciphering Raw Data in Neuro-Symbolic Learning with Provable Guarantees
Authors: Lue Tao, Yu-Xuan Huang, Wang-Zhou Dai, Yuan Jiang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on benchmark tasks with different knowledge bases validate the utility of the criterion. ... In this section, we conduct comprehensive experiments to validate the utility of the proposed criterion on various tasks. ... Table 1 presents the test performance of multi-layer perception (MLP) produced by hybrid learning methods on various datasets for the Conj Eq and Conjunction tasks. |
| Researcher Affiliation | Academia | Lue Tao,1,2 Yu-Xuan Huang,1,2 Wang-Zhou Dai,1,3 Yuan Jiang1,2 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China 3School of Intelligence Science and Technology, Nanjing University, China {taol, huangyx, daiwz, jiangy}@lamda.nju.edu.cn |
| Pseudocode | No | The paper describes the overall pipeline of algorithms in Appendix A using text and a figure, but it does not contain a formal pseudocode block or an explicitly labeled algorithm. |
| Open Source Code | Yes | The code is available for download.2 2https://github.com/Abductive Learning/ABL-TL |
| Open Datasets | Yes | Following previous work (Huang et al. 2021; Cai et al. 2021), we collect training sequences for the tasks by representing the handwritten symbols using instances from benchmark datasets including MNIST (Le Cun et al. 1998), EMNIST (Cohen et al. 2017), USPS (Hull 1994), KUZUSHIJI (Clanuwat et al. 2018), and FASHION (Xiao, Rasul, and Vollgraf 2017). |
| Dataset Splits | No | The paper states 'We random sample 6000 training images from each digit class of MNIST' and 'We randomly choose 1000 images from each digit class of MNIST to form the test set,' indicating some details for train and test splits for MNIST. However, it does not explicitly provide information on validation splits or specific numerical details for train/test/validation splits for all other datasets used (EMNIST, USPS, KUZUSHIJI, FASHION), stating only that they follow 'the same procedures'. |
| Hardware Specification | Yes | All experiments are repeated six times on Ge Force RTX 3090 GPUs, with the mean accuracy and standard deviation reported. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and MLP models, but it does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | All models are trained with the Adam optimizer (Kingma and Ba 2014) for 200 epochs, with an initial learning rate of 1e-3, which is then decayed by 0.1 at epochs 100 and 150. The mini-batch size is set to 256. |