A New Distribution on the Simplex with Auto-Encoding Applications

Authors: Andrew Stirn, Tony Jebara, David Knowles

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the distribution s utility in a variety of semi-supervised auto-encoding tasks. In all cases, the resulting models achieve competitive performance commensurate with their simplicity, use of explicit probability models, and abstinence from adversarial training.
Researcher Affiliation Collaboration Department of Computer Science Columbia University New York, NY 10027 {andrew.stirn,jebara,daknowles}@cs.columbia.edu jointly affiliated with New York Genome Center jointly affiliated with Spotify Technology S.A. jointly affiliated with Columbia University s Data Science Institute and the New York Genome Center
Pseudocode Yes Algorithm 1 A Generalized Stick-Breaking Process
Open Source Code Yes Our source code can be found at https://github.com/astirn/MV-Kumaraswamy.
Open Datasets Yes We utilize the Tensor Flow Datasets API, from which we source our data. For all experiments, we split our data into 4 subsets: unlabeled training (U) data, labeled training (L) data, validation data, and test data. For MNIST: |U| = 49, 400, |L| = 600, |validation| = |test| = 10, 0000. For SVHN: |U| = 62, 257, |L| = 1000, |validation| = 10, 000, |test| = 26, 032.
Dataset Splits Yes For all experiments, we split our data into 4 subsets: unlabeled training (U) data, labeled training (L) data, validation data, and test data. For MNIST: |U| = 49, 400, |L| = 600, |validation| = |test| = 10, 0000. For SVHN: |U| = 62, 257, |L| = 1000, |validation| = 10, 000, |test| = 26, 032.
Hardware Specification No The paper states 'We utilized GPU acceleration and found that cards with 8 GB of memory were sufficient,' but does not specify the exact GPU models (e.g., NVIDIA A100, RTX 2080 Ti) or CPU details used for the experiments.
Software Dependencies No The paper mentions 'Tensor Flow' and 'ADAM' but does not provide specific version numbers for these software components.
Experiment Setup Yes Our models are implemented in Tensor Flow and were trained using ADAM with a batch size B = 250 and 5 Monte-Carlo samples for each training example. We use learning rates 1 10 3 and 1 10 4 respectively for MNIST and SVHN. Other optimizer parameters were kept at Tensor Flow defaults.