Dr. Irina Higgins
Dr. Irina Higgins
UK
DeepMind
Senior Research Scientist
Dr. Irina Higgins is a Senior Research Scientist at DeepMind, where she aims to bring together insights from neuroscience, machine learning and physics to advance artificial intelligence.

Before DeepMind, Dr. Higgins did a DPhil in Computational Neuroscience at Oxford, worked on developing poker AI, and interned in the finance sector and Google Research.


Topic & Abstract

Unsupervised representation learning

Despite the advances in modern deep learning approaches, we are still quite far from the generality, robustness and data efficiency of biological intelligence. In my talk I will suggest that this gap may be narrowed by focusing on explicit unsupervised representation learning. This is in contrast to the implicit representation learning prevalent in modern end-to-end deep learning approaches. In particular, my talk will highlight the value of disentangled representations acquired in an unsupervised manner using methods inspired by biological intelligence. I will demonstrate how such representations enable the acquisition of reinforcement learning policies that are more robust to transfer scenarios compared to the standard approaches; and how they can serve as the foundation for learning abstract compositional visual concepts which underlie the ability to imagine meaningful and diverse samples beyond the training data distribution.

EVENTS IN YOUR AGENDA: 0