Research
I take a geometric perspective on optimisation, neural network architectures and statistical learning. I've also done work on geometry for its own sake, and maintain an interest in the geometry and topology of singular foliations.
Geometry of optimisation
I have worked on the geometric underpinnings of overparametrisation, and more recently on the geometry of loss landscapes and how they enable the training of neural networks with large-step gradient descent. This work brings deep learning theory closer to deep learning practice.
Geometry of architecture
Convolution algebras from noncommutative geometry show up in convolutional neural networks. My work develops a theory of convolutional neural networks over arbitrary finite-dimensional Lie groups. This work connects convolution algebras with architectures that respect transformations of their inputs.
Geometry of statistical learning
I have brought geometry to bear on statistical learning, introducing a new perspective on Wasserstein distances and average-case bounds based on a generalised notion of disintegration. I am also interested in how the geometry and implicit bias of learning algorithms affect generalisation.
Geometry of foliations
Holonomy groupoids, characteristic classes and equivariant KK-theory. My doctoral and subsequent work studies geometric representatives of invariants for regular and certain singular foliations.