Training Variational Autoencoders with Buffered Stochastic Variational Inference
This work addresses the suboptimal variational parameters in VAEs, an incremental improvement for machine learning practitioners using deep latent variable models.
The paper tackles the amortization gap in variational autoencoders by proposing Buffered Stochastic Variational Inference (BSVI), a refinement procedure that uses intermediate distributions and importance weights to create a new lower bound, resulting in consistent empirical outperformance over SVI.
The recognition network in deep latent variable models such as variational autoencoders (VAEs) relies on amortized inference for efficient posterior approximation that can scale up to large datasets. However, this technique has also been demonstrated to select suboptimal variational parameters, often resulting in considerable additional error called the amortization gap. To close the amortization gap and improve the training of the generative model, recent works have introduced an additional refinement step that applies stochastic variational inference (SVI) to improve upon the variational parameters returned by the amortized inference model. In this paper, we propose the Buffered Stochastic Variational Inference (BSVI), a new refinement procedure that makes use of SVI's sequence of intermediate variational proposal distributions and their corresponding importance weights to construct a new generalized importance-weighted lower bound. We demonstrate empirically that training the variational autoencoders with BSVI consistently out-performs SVI, yielding an improved training procedure for VAEs.