Infinite ShapeOdds: Nonparametric Bayesian Models for Shape Representations

Authors: Wei Xing, Shireen Elhabian, Robert Kirby, Ross T. Whitaker, Shandian Zhe6462-6469

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On synthetic and real-world data, we show the advantage of our method in both representation learning and latent structure discovery. Experiments Evaluating Compact Shape Representations. We first evaluated the quality of the learned representations by our method in the tasks of reconstruction and shape interpolation. Datasets. For a fair comparison, we used the same benchmark datasets as in the Shape Odds paper (Elhabian and Whitaker 2017): the Weizmann horse dataset (Borenstein and Ullman 2008), and the Caltech-101 motorcycle dataset (Fei-Fei, Fergus, and Perona 2007).
Researcher Affiliation Academia Wei Xing,1 Shireen Elhabian,1 Robert M. Kirby,1,2 Ross T. Whitaker,1,2 Shandian Zhe2 1Scientific Computing and Imaging Institute, University of Utah 2School of Computing, University of Utah {wxing, shireen, kirby, whitaker}@sci.utah.edu, zhe@cs.utah.edu
Pseudocode No The paper describes algorithms such as the truncated variational expectation-maximization algorithm, but it does not include a formal pseudocode block or algorithm box.
Open Source Code No The paper does not contain any statement about making its source code available or provide a link to a code repository for the methodology described.
Open Datasets Yes For a fair comparison, we used the same benchmark datasets as in the Shape Odds paper (Elhabian and Whitaker 2017): the Weizmann horse dataset (Borenstein and Ullman 2008), and the Caltech-101 motorcycle dataset (Fei-Fei, Fergus, and Perona 2007). Next, we examined our method on a real-world dataset, fashion MNIST (https://github.com/zalandoresearch/fashion-mnist).
Dataset Splits Yes To ensure the best performance for VAEs, we held 20% of the training data for validation. We varied the portion of the training images from {35%, 55%, 75%} and reconstructed the remaining images.
Hardware Specification No The paper does not provide specific details regarding the hardware used for running the experiments (e.g., GPU/CPU models, memory specifications).
Software Dependencies No The paper mentions software components like 'Adam' optimizer and 'Re LU activation function' but does not specify any version numbers for these or other software dependencies, such as libraries or frameworks.
Experiment Setup Yes For our method and GPLVM, we used the SE-ARD kernel and ran a maximum of 300 iterations for model estimation. Both methods were based on the L-BFGS optimization algorithm. We applied a Re LU activation function for all the hidden layers, except for the last layer of the decoder, for which we used a Sigmoid activation function to match the binary outputs (i.e., pixels). The Adam (Kingma and Ba 2014) algorithm was used for training. After running 1000 epochs, the model with the lowest validation error was chosen for testing. We set the truncation level for the DPM prior to 10.