Variational Generative Stochastic Networks with Collaborative Shaping

Authors: Philip Bachman, Doina Precup

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present empirical results on the MNIST and TFD datasets which show that our approach offers state-of-the-art performance, both quantitatively and from a qualitative point of view.
Researcher Affiliation Academia Mc Gill University, School of Computer Science
Pseudocode Yes Algorithm 1 Walkback for a General GSN
Open Source Code Yes Code implementing the models described in this paper is available online at: github.com/Philip-Bachman/ICML-2015.
Open Datasets Yes We present tests examining the behavior of our models on the MNIST and TFD datasets.
Dataset Splits Yes We performed model updates using gradients computed from mini-batches of 100 distinct examples from the training set... and ...150 examples selected at random from a validation set.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for the experiments.
Software Dependencies No The paper states 'We implemented our models in Python using the THEANO library (Bergstra et al., 2010)' but does not specify version numbers for Python or THEANO.
Experiment Setup Yes We represented qφ and p using neural networks with two hidden layers of 1500 rectified-linear units and we set the latent space Z to R64... We used a Maxout network (Goodfellow et al., 2013) with two hidden layers of 1200 units in 300 groups of 4 for the guide function f . For Lg in Eq. 7, we used a halfrectified elastic-net (Zou & Hastie, 2005), with the linear and quadratic parts weighted equally. We unrolled our chains for 6 steps... We performed model updates using gradients computed from mini-batches of 100 distinct examples from the training set... We trained using plain SGD. We pre-trained p and qφ as a variational autoencoder (VAE) for 100k updates...