Title: efficient-mixture-project
DNr: Berzelius-2024-126
Project Type: LiU Berzelius
Principal Investigator: Alexandra Hotti <hotti@kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2024-03-20 – 2024-05-01
Classification: 10201


In our project, we aim to significantly enhance density estimation in variational inference. We propose a model and an objective estimation technique that innovatively addresses the computational challenges of scaling up the number of mixture components, offering a scalable solution with minimal parameter increase and reduced inference time. By experimenting on the MNIST, Fashion-MNIST, CIFAR-10, and 8 different phylogenetic datasets, we anticipate groundbreaking results that could redefine scalability and efficiency in mixtures in variational inference and, in particular, variational autoencoders, its deep learning counterpart. This research promises not only theoretical advancements but also practical improvements in fields requiring complex density estimations.