Bidirectional Helmholtz Machines

Authors: Jorg Bornschein, Samira Shabanian, Asja Fischer, Yoshua Bengio

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we present experimental results obtained when applying the algorithm to various binary datasets.
Researcher Affiliation Academia J org Bornschein BORNJ@IRO.UMONTREAL.CA Samira Shabanian SHABANIS@IRO.UMONTREAL.CA Asja Fischer FISCHER@IRO.UMONTREAL.CA Yoshua Bengio BENGIOY@IRO.UMONTREAL.CA Dept. Computer Science and Operations Research, University of Montreal Canadian Institute for Advanced Research (CIFAR)
Pseudocode Yes Algorithm 1 Training p (x) using K importance samples
Open Source Code Yes Our implementation is available at https://github.com/jbornschein/bihm.
Open Datasets Yes We use the MNIST dataset that was binarized according to (Murray & Salakhutdinov, 2009) and which we downloaded in binarized form (Larochelle, 2011).
Dataset Splits No The paper mentions training and testing but does not explicitly detail a separate validation dataset split with specific percentages or counts.
Hardware Specification Yes Estimating based on 10 million samples takes less than 2 minutes on a GTX980 GPU.
Software Dependencies No The paper mentions software like Theano and Blocks with citations, but does not specify their version numbers for reproducibility.
Experiment Setup Yes We train all models using Adam (Kingma & Ba, 2014) with a mini-batch size of 100. We initialize the weights according to Glorot & Bengio (2010), set the biases to -1, and use L1 regularization λ=10 3 on all the weights. Our implementation is available at https://github.com/jbornschein/bihm. We use a learning rate of 10 2 or 10 3 for all the experiments.