Hermanni Hälvä

I am a machine learning PhD student at the University of Helsinki, supervised by Aapo Hyvärinen. My thesis focuses on nonlinear ICA and identifiable representation learning. Before the PhD, I completed a masters degree in computational statistics and machine learning at UCL.

Email  /  GitHub  /  Google Scholar

Research

My main areas of interest are deep generative models and representation learning. I am especially interested in disentanglement learning via nonlinear ICA and the identifiability of these type of models. I am also broadly interested in probabilistic methods, approximate inference and computer vision.

Identifiable Feature Learning for Spatial Data with Nonlinear ICA

A new algorithm that allows nonlinear ICA to be applied to data with arbitrary high-dimensional latent structures. This model relies on t-process latent space, and was developed with spatio-temporal data in mind.

Disentangling Identifiable Features from Noisy Data with Structured Nonlinear ICA

General framework for identifiable disentanglement on structured data with nonlinear ICA. Identifiability holds even in the presence of uknown and arbitrary output noise.

Independent Innovation Analysis for Nonlinear Vector Autoregressive Process

Nonlinear vector autoregressive process where the innovations at each time point are assumed indepndent and are proven to be identifiable. Can be interpreted as nonlinear ICA for autoregressive time-series.

Hidden Markov Nonlinear ICA: Unsupervised Learning from Nonstationary Time Series

We show that temporal nonstationary as represented by hidden markov chain identifies nonlinear ICA.

Based on Leonid Keselman's Jekyll Fork of Jon Barron's website.