Mind the Gap when Conditioning Amortised Inference in Sequential\n Latent-Variable Models

2021/01/18 by Justin Bayer, Maximilian Soelch, Bayer, Justin +7 · 1 citation
Computer Science · #FOS: Computer and information sciences #Gaussian Processes and Bayesian Inference #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning in Healthcare

paper · pdf · doi:10.48550/arxiv.2101.07046

openalex publication_date 2021/01/18 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

Amortised inference enables scalable learning of sequential latent-variable\nmodels (LVMs) with the evidence lower bound (ELBO). In this setting,\nvariational posteriors are often only partially conditioned. While the true\nposteriors depend, e.g., on the entire sequence of observations, approximate\nposteriors are only informed by past observations. This mimics the Bayesian\nfilter -- a mixture of smoothing posteriors. Yet, we show that the ELBO\nobjective forces partially-conditioned amortised posteriors to approximate\nproducts of smoothing posteriors instead. Consequently, the learned generative\nmodel is compromised. We demonstrate these theoretical findings in three\nscenarios: traffic flow, handwritten digits, and aerial vehicle dynamics. Using\nfully-conditioned approximate posteriors, performance improves in terms of\ngenerative modelling and multi-step prediction.\n

Citations

• A Recurrent Latent Variable Model for Sequential Data
• Auto-Encoding Variational Bayes
• Training Products of Experts by Minimizing Contrastive Divergence
• Bidirectional recurrent neural networks
• Deep Bidirectional and Unidirectional LSTM Recurrent Neural Network for Network-wide Traffic Speed Prediction
• Traffic Graph Convolutional Recurrent Neural Network: A Deep Learning Framework for Network-Scale Traffic Learning and Forecasting
• Approximate Bayesian inference in spatial environments
• Variational Tracking and Prediction with Generative Disentangled\n State-Space Models
• Learning to Fly via Deep Model-Based Reinforcement Learning

Cited by

• Latent Matters: Learning Deep State-Space Models

Related

• Amortizing intractable inference in diffusion models for vision, language, and control
• A Unified Theory of Exact Inference and Learning in Exponential Family Latent Variable Models
• Amortized Variational Inference: When and Why?
• Iterative Amortized Inference
• Adversarial robustness of amortized Bayesian inference