MAGANet: Achieving Combinatorial Generalization by Modeling a Group Action

Authors: Geonho Hwang, Jaewoong Choi, Hyunsoo Cho, Myungjoo Kang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To test the combinatorial generalizability, we evaluated our model in two settings: Recombination-to-Element and Recombination-to-Range. The experiments demonstrated that our method has quantitatively and qualitatively superior generalizability and generates better images than traditional models.
Researcher Affiliation Academia 1Korea Institute for Advanced Study, Seoul, South Korea 2Seoul National University, Seoul, South Korea.
Pseudocode Yes Detailed architectures are in Appendix A. Table 4. The Encoder Architecture. Table 5. The Flow Step Architecture. Table 6. The Decoder Architecture. Table 7. The Flow Net Architecture.
Open Source Code No The paper does not provide any explicit statements about the release of source code or links to a code repository.
Open Datasets Yes We evaluate our model on two datasets, the d Sprites dataset (Matthey et al., 2017) and the 3D Shapes dataset (Burgess & Kim, 2018).
Dataset Splits Yes To test the combinatorial generalization property of the model, we measure the reconstruction error while separating the training and test data. Following the evaluation protocol of Montero et al. (2020), we evaluate the combinatorial generalization under two settings, the Recombination-to Element and the Recombination-to-Range.
Hardware Specification No The paper does not specify any hardware used for running the experiments (e.g., CPU, GPU models, or cloud computing instances with their specifications).
Software Dependencies No The paper mentions 'The optimizer is Adam' but does not provide specific version numbers for Adam or any other software libraries or dependencies used.
Experiment Setup Yes The optimizer is Adam, with a learning rate of 0.0005. The dimension of the latent variable is set to 10 and the batch size is 64. For the regularizing coefficients β, we set the value to βrecon latent = 300, and βKL = 300. We trained 100 epochs for both datasets three times and took the model with the best binary cross entropy loss model.