Modern single-cell datasets often are often extremely high-dimensional which presents challenges for traditional methods of data analysis. However, such data sets often have a hidden, low-dimensional underlying structure where, for example, gene coordinate and redundancy may cause the data to lie upon a comparatively low-dimensional manifold. This motivates methods which find and/or utilize this hidden structure.
My talk will focus on methods from geometric deep learning which approximate the underlying data manifold via a graph and use wavelet-based graph neural networks, based on the geometric scattering transform. We show that these methods capture the underlying data geometry and are able to capture both the local and global structure of the data.
Copyright. All Rights Reserved.