Multi-sample, importance-weighted variational autoencoders (IWAE) give tighter bounds and more accurate uncertainty estimates than variational autoencoders (VAEs) trained with a standard single-sample objective. However, IWAEs scale poorly: as the latent dimensionality grows, they require exponentially many samples to retain the benefits of importance weighting. While sequential Monte-Carlo (SMC) can address this problem, it is prohibitively slow because the resampling step imposes sequential structure which cannot be parallelised, and moreover, resampling is non-differentiable which is problematic when learning approximate posteriors. To address these issues, we developed tensor Monte-Carlo (TMC) which gives exponentially many importance samples by separately drawing $K$ samples for each of the $n$ latent variables, then averaging over all $K^n$ possible combinations. While the sum over exponentially many terms might seem to be intractable, in many cases it can be computed efficiently as a series of tensor inner-products. We show that TMC is superior to IWAE on a generative model with multiple stochastic layers trained on the MNIST handwritten digit database, and we show that TMC can be combined with standard variance reduction techniques.

Summary: The authors describe an improved objective function for variational inference. In the spirit of the Importance Weighted Autoencoder they use multiple samples from the approximating distribution to obtain a tighter bound on the log marginal probability. The key insight of the paper is that they can use all combinations (across subsets of parameters) of samples drawn from the approximating distribution to compute marginal probability estimator. Naively this would require computation that scales exponentially in the number of subsets. The authors reduce this complexity by exploiting dependency structures in the generative model.

Multi-sample, importance-weighted variational autoencoders (IWAE) give tighter bounds and more accurate uncertainty estimates than variational autoencoders (VAEs) trained with a standard single-sample objective. However, IWAEs scale poorly: as the latent dimensionality grows, they require exponentially many samples to retain the benefits of importance weighting. While sequential Monte-Carlo (SMC) can address this problem, it is prohibitively slow because the resampling step imposes sequential structure which cannot be parallelised, and moreover, resampling is non-differentiable which is problematic when learning approximate posteriors. To address these issues, we developed tensor Monte-Carlo (TMC) which gives exponentially many importance samples by separately drawing K samples for each of the n latent variables, then averaging over all K n possible combinations. While the sum over exponentially many terms might seem to be intractable, in many cases it can be computed efficiently as a series of tensor inner-products. We show that TMC is superior to IWAE on a generative model with multiple stochastic layers trained on the MNIST handwritten digit database, and we show that TMC can be combined with standard variance reduction techniques.

Multi-sample, importance-weighted variational autoencoders (IWAE) give tighter bounds and more accurate uncertainty estimates than variational autoencoders (VAEs) trained with a standard single-sample objective. However, IWAEs scale poorly: as the latent dimensionality grows, they require exponentially many samples to retain the benefits of importance weighting. While sequential Monte-Carlo (SMC) can address this problem, it is prohibitively slow because the resampling step imposes sequential structure which cannot be parallelised, and moreover, resampling is non-differentiable which is problematic when learning approximate posteriors. To address these issues, we developed tensor Monte-Carlo (TMC) which gives exponentially many importance samples by separately drawing $K$ samples for each of the $n$ latent variables, then averaging over all $K n$ possible combinations. While the sum over exponentially many terms might seem to be intractable, in many cases it can be computed efficiently as a series of tensor inner-products.

Multi-sample objectives improve over single-sample estimates by giving tighter variational bounds and more accurate estimates of posterior uncertainty. However, these multi-sample techniques scale poorly, in the sense that the number of samples required to maintain the same quality of posterior approximation scales exponentially in the number of latent dimensions. One approach to addressing these issues is sequential Monte Carlo (SMC). However for many problems SMC is prohibitively slow because the resampling steps imposes an inherently sequential structure on the computation, which is difficult to effectively parallelise on GPU hardware. We developed tensor Monte-Carlo to address these issues. In particular, whereas the usual multi-sample objective draws $K$ samples from a joint distribution over all latent variables, we draw $K$ samples for each of the $n$ individual latent variables, and form our bound by averaging over all $K^n$ combinations of samples from each individual latent. While this sum over exponentially many terms might seem to be intractable, in many cases it can be efficiently computed by exploiting conditional independence structure. In particular, we generalise and simplify classical algorithms such as message passing by noting that these sums can be computed can be written in an extremely simple, general form: a series of tensor inner-products which can be depicted graphically as reductions of a factor graph. As such, we can straightforwardly combine summation over discrete variables with importance sampling over importance sampling over continuous variables.