Learning high-level, information-rich, and robust representations from data is an important research area that can help to solve many downstream tasks in machine learning.

In conditions where the label information used to train the models is imprecise, or the models face distribution shift in testing data, the robustness of the learned representations plays an important role in the performance and generalisation of such models.

In this talk, 3 examples of such conditions will be discussed: 1) learning under distribution mismatch between training and testing, 2) learning from weakly labeled data, and 3) learning to generate with high diversity without labels. I will show how deep neural networks can be leveraged to learn high-level and robust representations from acoustic recordings, long genomic sequences, and images, under the above conditions. And I demonstrate how end-to-end representation learning can help to tackle several tasks in Computational Auditory Scene Analysis, Structural Variations/Copy Number Variations annotations of genomic data, and image synthesis.

>Link to the announcement, opens an external URL in a new window

>Link to the archive, opens an external URL in a new window