The foundation of modern artificial intelligence is machine learning: Intelligence emerges from the analysis of large amounts of input data. An important goal of machine learning is to find underlying factors or causes in data, which is a special case of unsupervised learning. This project is about a specific model for unsupervised learning, called independent component analysis (ICA). While the model is well-known in the linear case, finding general, nonlinear independent components is a very challenging problem and little progress was made until very recently. This project attempts to take that theoretical framework and make it generally applicable. We need to develop the theory of such nonlinear ICA, design new algorithms, and explore different cases where the method could be used for real data. The goal is to make nonlinear ICA the dominant paradigm in unsupervised learning, and in particular for learning in neural networks, which is also called deep learning.

2020

2024