Relational Learning with Variational Bayes
Authors: Kuang-Hung Liu
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experimental results from applying VRL to a variety of relational learning tasks. |
| Researcher Affiliation | Industry | Kuang-Hung Liu Exxon Mobil Research and Engineering kuang-hung.liu@exxonmobil.com |
| Pseudocode | Yes | Algorithm 1 VRL with RPDA |
| Open Source Code | Yes | Full implementation details are provided in Appendix E.1 and source code for reproducible results is available online1. 1https://github.com/kh1iu/vrl.git |
| Open Datasets | Yes | MNIST dataset (available under CC BY-SA 3.0 license) (Le Cun & Cortes, 2010). Omniglot dataset (available under MIT license) (Lake et al., 2015). Yale Face Database (see (Georghiades, 2001) for dataset permission) (Belhumeur et al., 1997). Extended Yale Face Database B (see (Georghiades, 2001) for dataset permission) (Georghiades et al., 2001). Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) (available under CC BY-NC-SA 4.0 license) (Livingstone & Russo, 2018). |
| Dataset Splits | Yes | Omniglot dataset... is split into a background (training) set of 30 alphabets and an evaluation (testing) set of 20 alphabets (Lake et al., 2015). and To quantitatively evaluate the clustering performance, we calculate the classification error rate of a hold-out dataset based on a simple classifier that was trained on 5% of the data with label. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used for experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and 'Tensorlayer' but does not specify their version numbers or other key software dependencies with explicit version information needed for full reproducibility. |
| Experiment Setup | Yes | All MLPs with parameters θ and φ were jointly trained for 300k iterations (without batchnormalization, weight decay, nor dropout) to maximize e L(i) RPDA in Eq. (4) with using Adam optimizer (learning rate=0.0004, β1=0.9, β1=0.999) (Kingma & Ba, 2015). Minibatches of size M=100 were used. We anneal the learning rate (0.0004 base learning rate) with step decay (factor of 0.5 every 100k iterations). When Gumbel-Softmax distributions is used, we anneal the softmax temperature τ from 1.0 to 0.5 with exponential decay (decay rate=0.00005). |