Coupled Variational Autoencoder
Authors: Xiaoran Hao, Patrick Shafto
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Simulations on synthetic and real data show that the C-VAE outperforms alternatives including VAE, WAE, and Info VAE in fidelity to the data, quality of the latent representation, and in quality of generated samples. In this section, we explore various properties of C-VAE on a selection of synthetic and real datasets. And We also compare it with other autoencoder models including VAE, VAENF2 and Info VAE. |
| Researcher Affiliation | Academia | 1Department of Mathematics and Computer Science, Rutgers University-Newark, New Jersey, USA 2School of Mathematics, Institute for Advanced Study, New Jersey, USA. Correspondence to: Xiaoran Hao <xh197@rutgers.edu>. |
| Pseudocode | Yes | Algorithm 1 Primal Strategy and Algorithm 2 Dual Strategy |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the source code for the described methodology, nor does it provide any links to a code repository. |
| Open Datasets | Yes | MNIST. We then run our model on a subset of MNIST which consists of 2560 different hand-written digits images from 10 classes. |
| Dataset Splits | No | The paper describes the datasets used ('Mixture of 25 Gaussians', 'MNIST') and sample counts for the synthetic data, but does not explicitly provide details about training, validation, or test splits for either dataset. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments, only general statements like 'Models are trained through Adam optimizer'. |
| Software Dependencies | No | The paper mentions using 'Re LU activations', 'Adam optimizer', and 'convolutional layers' but does not specify any software libraries or frameworks with their respective version numbers (e.g., PyTorch 1.9, TensorFlow 2.x). |
| Experiment Setup | Yes | For all encoder and decoder neural networks, we used four fully connected layers with Re LU activations with Batch normalizations in between. All hidden layers have 256 neurons and we choose the latent dimension dz = 2 to plot the latent space. The prior is a 2D Gaussian with a mean of 0 and a standard deviation of 1, i.e., p(z) = N(0, Id). We pick the best model based on hyperparameter searching. Latent dimension dz = 2. All the autoencoder networks have the same architecture. We use convolutional layers paired with Re LU as the building block for the encoding/decoding networks. We choose Bernoulli likelihood in the experiments. Models are trained through Adam optimizer with β1 = 0.9 and β2 = 0.999. |