Unsupervised Object Representation Learning using Translation and Rotation Group Equivariant VAE

Authors: Alireza Nasiri, Tristan Bepler

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

Reproducibility Variable Result LLM Response
Research Type Experimental In comprehensive experiments, we show that TARGET-VAE learns disentangled representations without supervision that significantly improve upon, and avoid the pathologies of, previous methods.
Researcher Affiliation Academia Alireza Nasiri Simons Machine Learning Center New York Structural Biology Center anasiri@nysbc.org Tristan Bepler Simons Machine Learning Center New York Structural Biology Center tbepler@nysbc.org
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Our code is available at https://github.com/SMLC-NYSBC/TARGET-VAE.
Open Datasets Yes Datasets We use two variants of the MNIST dataset: 1) (MNIST(N)) MNIST digits... and 2) (MNIST(U)) MNIST digits... (Appendix Figure 1). ... EMPIAR-10025 [34]... EMPIAR-10029 is a simulated EM dataset of Gro EL particles.
Dataset Splits Yes In both datasets, the train and test sets have 60,000, and 10,000 images of dimensions 50x50 pixels, respectively. ... We randomly select 10% of the images as the validation set to monitor the training process.
Hardware Specification Yes We run all our experiments in a single NVIDIA A100 GPU, with 80 GB memory.
Software Dependencies No The paper mentions specific optimizers and techniques used (e.g., 'ADAM optimizer', 'Gumbel-Softmax') but does not specify version numbers for any programming languages, libraries, or frameworks used in the implementation.
Experiment Setup Yes We use leaky-ReLU activation functions, and the batch size for all the experiments is set to 100. We use ADAM optimizer [30], with learning rate of 2e-4, and the learning rate is decayed by the factor of 0.5 after no improvements in the loss for 10 epochs. We run the training for a maximum of 500 epochs with early-stopping in case of no improvements in the loss for 20 epochs.