Unsupervised Learning of Equivariant Structure from Sequences

Authors: Takeru Miyato, Masanori Koyama, Kenji Fukumizu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted several experiments to investigate the efficacy of our framework. In this section, we briefly explain the experimental settings. For more details, please see Appendix D. We tested our framework on Sequential MNIST, 3DShapes [5], and Small NORB [48].
Researcher Affiliation Collaboration Takeru Miyato 1,2 Masanori Koyama 1 Kenji Fukumizu3,1 equal contribution 1Preferred Networks, Inc. 2University of Tübingen 3The Institute of Statistical Mathematics
Pseudocode No The paper includes a diagram (Figure 2) to illustrate the procedure but does not contain a formal pseudocode or algorithm block.
Open Source Code Yes The code is available at https://github.com/takerum/meta_sequential_prediction.
Open Datasets Yes We tested our framework on Sequential MNIST, 3DShapes [5], and Small NORB [48]. Sequential MNIST is created from MNIST dataset [47].
Dataset Splits No The paper mentions that it uses 'training and test sets' (Section D) and an internal split (sc, sp) for each sequence (Section 4), but it does not provide specific percentages, absolute sample counts, or clear predefined split references for the overall train/validation/test dataset splits.
Hardware Specification Yes We used one NVIDIA A100-80GB GPU to train each model.
Software Dependencies No The paper mentions optimizers (Adam) and architectures (ResNet), but it does not specify any software libraries or frameworks with their version numbers (e.g., PyTorch version, Python version, specific package versions) required to reproduce the experiments.
Experiment Setup Yes We trained our models using Adam [37] for 500 epochs with a learning rate of 1e-4. The batch size was set to 128. We used Layer Normalization [2] in our ResNet architecture and used weight standardization [55] for convolution layers.