Mind the Gap when Conditioning Amortised Inference in Sequential Latent-Variable Models
This addresses a critical issue in scalable learning of sequential latent-variable models for researchers and practitioners, but it is incremental as it builds on existing amortised inference and ELBO frameworks.
The paper tackled the problem of partially-conditioned amortised inference in sequential latent-variable models, showing that it leads to compromised generative models by approximating products of smoothing posteriors instead of the true Bayesian filter. Using fully-conditioned approximate posteriors improved performance in generative modeling and multi-step prediction across traffic flow, handwritten digits, and aerial vehicle dynamics scenarios.
Amortised inference enables scalable learning of sequential latent-variable models (LVMs) with the evidence lower bound (ELBO). In this setting, variational posteriors are often only partially conditioned. While the true posteriors depend, e.g., on the entire sequence of observations, approximate posteriors are only informed by past observations. This mimics the Bayesian filter -- a mixture of smoothing posteriors. Yet, we show that the ELBO objective forces partially-conditioned amortised posteriors to approximate products of smoothing posteriors instead. Consequently, the learned generative model is compromised. We demonstrate these theoretical findings in three scenarios: traffic flow, handwritten digits, and aerial vehicle dynamics. Using fully-conditioned approximate posteriors, performance improves in terms of generative modelling and multi-step prediction.