Optimal Binary Autoencoding with Pairwise Correlations

Authors: Akshay Balsubramani

ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments in Section 6 show extremely competitive results with equivalent fully-connected autoencoders trained with backpropagation. The datasets we use are first normalized to [0, 1], and then binarized by sampling each pixel stochastically in proportion to its intensity, following prior work (Salakhutdinov & Murray (2008)).
Researcher Affiliation Academia Akshay Balsubramani Stanford University abalsubr@stanford.edu
Pseudocode Yes Algorithm 1 Pairwise Correlation Autoencoder (PC-AE) Input: Size-n dataset ˆX, number of epochs T
Open Source Code Yes TensorFlow code available at https://github.com/aikanor/pc-autoencoder .
Open Datasets Yes The datasets we use are first normalized to [0, 1], and then binarized by sampling each pixel stochastically in proportion to its intensity, following prior work (Salakhutdinov & Murray (2008)). We use the preprocessed version of the Omniglot dataset found in Burda et al. (2016), split 1 of the Caltech-101 Silhouettes dataset, the small not MNIST dataset, and the UCI Adult (a1a) dataset.
Dataset Splits No The paper mentions 'early stopping on the test set' and '10-fold cross-validation' for 'not MNIST', but it does not explicitly describe a separate validation dataset split with specific percentages or counts for all experiments.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as CPU/GPU models, memory, or cloud instance types.
Software Dependencies No The paper mentions 'TensorFlow code available at https://github.com/aikanor/pc-autoencoder', and the use of 'Adagrad (Duchi et al. (2011))' and the 'Adam method with default parameters (Kingma & Ba (2014))' for optimization, but it does not specify version numbers for these software components or libraries.
Experiment Setup Yes We used minibatches of size 250. All standard autoencoders use the Xavier initialization and trained for 500 epochs or using early stopping on the test set. We compare to a basic AE with a single hidden layer, trained using the Adam method with default parameters (Kingma & Ba (2014)).