Reweighted Wake-Sleep
arXiv:1406.2751
Abstract
Training deep directed graphical models with many hidden variables and performing inference remains a major challenge. Helmholtz machines and deep belief networks are such models, and the wake-sleep algorithm has been proposed to train them. The wake-sleep algorithm relies on training not just the directed generative model but also a conditional generative model (the inference network) that runs backward from visible to latent, estimating the posterior distribution of latent given visible. We propose a novel interpretation of the wake-sleep algorithm which suggests that better estimators of the gradient can be obtained by sampling latent variables multiple times from the inference network. This view is based on importance sampling as an estimator of the likelihood, with the approximate inference network as a proposal distribution. This interpretation is confirmed experimentally, showing that better likelihood can be achieved with this reweighted wake-sleep procedure. Based on this interpretation, we propose that a sigmoidal belief network is not sufficiently powerful for the layers of the inference network in order to recover a good estimator of the posterior distribution of latent variables. Our experiments show that using a more powerful layer model, such as NADE, yields substantially better generative models.
References in corpus (7)
- Auto-Encoding Variational Bayes
- Stochastic Backpropagation and Approximate Inference in Deep Generative Models
- Modeling Temporal Dependencies in High-Dimensional Sequences: Application to Polyphonic Music Generation and Transcription
- Better Mixing via Deep Representations
- Deep AutoRegressive Networks
- A Deep and Tractable Density Estimator
- Efficient Gradient-Based Inference through Transformations between Bayes Nets and Neural Nets
Cited by in corpus (23)
- Towards Biologically Plausible Deep Learning
- How Auto-Encoders Could Provide Credit Assignment in Deep Networks via Target Propagation
- Boundary-Seeking Generative Adversarial Networks
- Quantum-assisted Helmholtz machines: A quantum-classical deep learning framework for industrial datasets in near-term devices
- On the Origin of Deep Learning
- Neural Adaptive Sequential Monte Carlo
- Discrete Variational Autoencoders
- ZhuSuan: A Library for Bayesian Deep Learning
- Variational Memory Addressing in Generative Models
- Learning Wake-Sleep Recurrent Attention Models
- Deep Directed Generative Autoencoders
- Learning to Adapt by Minimizing Discrepancy
- Meta-Amortized Variational Inference and Learning
- Variational Autoencoders for Feature Detection of Magnetic Resonance Imaging Data
- Learning Deep Generative Models with Annealed Importance Sampling
- Joint Stochastic Approximation learning of Helmholtz Machines
- Filtering Variational Objectives
- Iterative Refinement of the Approximate Posterior for Directed Belief Networks
- On the Difficulty of Unbiased Alpha Divergence Minimization
- Joint Stochastic Approximation and Its Application to Learning Discrete Latent Variable Models
- Amortized Population Gibbs Samplers with Neural Sufficient Statistics
- GSNs : Generative Stochastic Networks
- Semi-supervised Sequential Generative Models