Unsupervised learning with deep generative models

Témavezető:	Orbán Gergő
	MTA Wigner FK, Komputációs tudományok osztálya
email:	orban.gergo@wigner.mta.hu

Projekt leírás

Machine learning has recently achieved breakthroughs by developing deep discriminative models (commonly termed deep learning models) that are capable of human-level learning about natural-complexity images. A drawback of these methods is that they require labeled data: for every single input image a desired output is required. In more general learning settings such labels are not available. A more recent advance in machine learning is the development of Variational Autoencoders, a model class that is performing learning in an unsupervised fashion. These models are attractive since they promise to discover the features underlying observations automatically, a topic called representation learning. In this research topic we aim to investigate the principles underlying learning hierarchies of features. In the proposed project we explore the principles of learning useful representations and/r how these representations can be used to flexibly learn novel tasks.

The research topic provides an opportunity to work on a hot topic in machine learning, to get familiar with topics such as probabilistic generative models, Bayesian methods, deep learning, and to get to know modern programming environments. It requires experience with modern scientific computing tools: python / pytorch / tensorflow and basic knowledge of the principles of unsupervised learning (PCA, Factor Analysis, Bayesian inference) are welcome.