Twisted variational sequential monte carlo

Latent variable models have achieved wide-ranging success in generative modeling of images, speech, and video. Recent work has developed specialized objectives for fitting sequential latent variable models [1, 2, 3], but these methods are based on filtering sequential Monte Carlo (SMC), which can prevent them from successfully modeling data with long-term dependencies. In this work, we introduce twisted variational sequential Monte Carlo (TVSMC), a family of variational objectives that use learned tilting functions to approximate smoothing SMC without direct access to the smoothing distributions. We demonstrate that models trained with TVSMC can outperform models trained with VSMC, and theoretically, the TVSMC bound can become tight.