Ljubljana, Slovenia

Deep Language Learning: Modeling language from Raw Speech

For the first time in history, we can model language from raw speech in a fully unsupervised manner. Using deep generative models trained on speech, we can model phonetic, phonological, morphological, and even basic syntactic and lexical semantic learning. This course will introduce students to deep learning and techniques to uncover linguistically meaningful representations in deep neural networks trained on raw speech. We will focus on convolutional neural networks, an architecture that is inspired by biological neural processing and has seen many applications in visual and audio processing. We will learn to train a generative adversarial network on spoken language, find disentangled causal structure in the hidden space of these networks, introspect their intermediate representations, and use these outputs for language modeling. Understanding how deep neural networks learn has consequences both for linguistic theory and cognitive science as well as for building interpretable machine learning.

Course leader

Gasper Begus

Target group

Students

Fee info

EUR 490: Early student registration
EUR 690: Early non-academic registration

Scholarships

There are several scholarship options which you can read about on our website.