ByPE-VAE: Bayesian Pseudocoresets Exemplar VAE

Authors: Qingzhong Ai, LIRONG HE, SHIYU LIU, Zenglin Xu

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that By PE-VAE can achieve competitive improvements over the state-of-the-art VAEs in the tasks of density estimation, representation learning, and generative data augmentation. Finally, we compare By PE-VAE with several state-of-the-art VAEs in a number of tasks, including density estimation, representation learning and generative data augmentation. Experimental results demonstrate the effectiveness of By PE-VAE on Dynamic MNIST, Fashion MNIST, CIFAR10, and Celeb A. 5 Experiments
Researcher Affiliation Academia Qingzhong Ai1 Lirong He1 Shiyu Liu1 Zenglin Xu2,3, 1School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu China 2School of Science and Technology, Harbin Institute of Technology Shenzhen, Shenzhen China 3Department of Network Intelligence, Peng Cheng National Lab, Shenzhen, China {qzai,lirong_he}@std.uestc.edu.cn, shyu.liu@foxmail.com, xuzenglin@hit.edu.cn
Pseudocode Yes Algorithm 2: The Optimization Algorithm for By PE-VAE
Open Source Code Yes Code is available at https://github.com/Aiqz/By PE-VAE.
Open Datasets Yes We evaluate the By PE-VAE on four datasets across several tasks based on multiple network architectures. Specifically, the tasks involve density estimation, representation learning and data augmentation, the used four datasets include MNIST, Fashion-MNIST, CIFAR10, and Celeb A, respectively.
Dataset Splits No No explicit training/validation/test dataset splits (e.g., percentages, sample counts) are provided for reproducibility beyond standard benchmark usage.
Hardware Specification Yes All experiments are run on a single Nvidia 1080Ti GPU.
Software Dependencies No No specific software dependencies with version numbers (e.g., library or framework versions like PyTorch 1.9, TensorFlow 2.x) are explicitly listed in the paper.
Experiment Setup Yes The ADAM algorithm with normalized gradients [23, 24, 25] is used for optimization and learning rate is set to 5e-4. And we use KL annealing for 100 epochs and early-stopping with a look ahead of 50 epochs.