Learning Symmetric Rules with SATNet

Authors: Sangho Lim, Eun-Gyeol Oh, Hongseok Yang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments with Sudoku and Rubik s cube show the substantial improvement of Sym SATNet over the baseline SATNet.
Researcher Affiliation Academia Sangho Lim School of Computing KAIST Daejeon, South Korea lim.sang@kaist.ac.kr Eun-Gyeol Oh Graduate School of Information Security KAIST Daejeon, South Korea eun-gyeol.oh@kaist.ac.kr Hongseok Yang School of Computing and Kim Jaechul Graduate School of AI, KAIST Discrete Mathematics Group, Institute for Basic Science (IBS) Daejeon, South Korea hongseok.yang@kaist.ac.kr
Pseudocode Yes Algorithm 1 SYMFIND with a threshold λ > 0
Open Source Code No The paper states 'Sym SATNet is implemented based on the SATNet code [26] available under the MIT License.' This refers to a third-party tool they used, not their own source code for Sym SATNet being made publicly available.
Open Datasets Yes We used 9K training and 1K test examples generated by the Sudoku generator [21]. [21] is Kyubyong Park. Can convolutional neural networks crack sudoku puzzles? https://github. com/Kyubyong/sudoku, 2018.
Dataset Splits Yes We used 8K training, 1K validation, and 1K test examples to train Sym SATNet with symmetries found by SYMFIND and the validation step.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions 'Adam optimizer' but does not provide specific software dependencies with version numbers (e.g., Python version, PyTorch/TensorFlow versions, specific library versions).
Experiment Setup Yes We used binary cross entropy loss and Adam optimizer [16], with the learning rate η = 2 10 3 for SATNet-Plain and SATNet-300aux as the original work and η = 4 10 2 for Sym SATNet. We trained Sym SATNet, SATNet-Plain, and SATNet-300aux for 100 epochs, under the same configuration as in the Sudoku case.