Скачать с ютуб Scalable probabilistic modeling and inference with structured latent representations в хорошем качестве

Scalable probabilistic modeling and inference with structured latent representations

Ph.D. thesis defense of Liam Kruse Modeling and inference tasks such as density estimation and importance sampling rely on probabilistic models to represent complex probability distributions. By exploiting favorable geometries in the latent spaces of probabilistic models, we can impose inductive biases that scale the training, learning, and inference processes. This work introduces three main contributions to address the challenges of scaling generative models and inference algorithms to high dimensions. First, we perform importance sampling in the latent space of pretrained normalizing flow models, finding that the simple latent distributions are easier to explore than the data distributions when fitting proposal densities. Second, we develop an approach to high-dimensional importance sampling using low-rank mixture proposals, which balance expressivity and computational efficiency by capturing dominant data subspaces with a latent variable model. Lastly, we move beyond importance sampling and safety validation to develop a framework for improving the normalizing flow training procedure. By replacing simple latent distributions with expressive low-rank mixture models, we effectively warm-start the flow training process, enabling flows to model complex distributions with less computational effort at training and inference time. We apply these algorithms to challenging high-dimensional tasks such as image generation and safety validation for a ground collision avoidance system. Slides available at: https://web.stanford.edu/group/sisl/p...