Geometric Data Analytics, Inc.

Expertise: Learning

Pipeline using TDA to generate features for learning.

Learning (Machine/Bayesian Learning) directly from a dataset is a popular, powerful way to discover information from data collected in almost any environment. The challenge, however, is to find the right set of “feature primitives” to work with in doing this learning. Even in deep-learning, which is often described as an unsupervised, meaning that information can be learned without an expert supervisor, one needs to have the right coordinates for the dataset. These themselves are best when they are feature primitives. For example, if you want to decide automatically what kind of instrument is playing a particular song, it makes sense to use features known to musicians (and engineers who study music) that capture the way a human ear perceives the music.

Geometry and Topology can be used to discover and code new feature primitives that derive directly from the shape of the dataset. These then provide the input to learning algorithms. Even deep-learning performs better on some problems with GDA/TDA features as input.