# Applications of Differential Geometry in Artificial Intelligence

I am new to this wonderful site. I searched around a bit but I couldn't find any well-discussed posts on applications of differential geometry to artificial intelligence, or more generally to computer science. I came across Riemannian Geometry a few months back via YouTube(lol) and have been hooked on it since. I even bought a textbook and started learning from there. I am an engineer and I have sufficient math skills to make sense of the book, but it hardly has any real life applications. Almost everything is "Prove this" or "Theorem that". So I wanted to know if there are any real-life implementable applications for CS. More specifically, in the field of AI and Machine Learning.

From what I have understood, differential geometry allows us to "see","understand" and "analyze" curves in higher dimensional spaces. Is this accurate? And can this help in AI and Machine Learning? In subtopics like Natural Language Processing, Robotics, Computer Vision, Data analysis? I would sure like to start off with a simple project which helps me understand differential geometry better.

Thanks

PS: Sorry if I have not framed this question well, it is my first question.

• I would say that the most frequent/mainstream use of differential geometry in CS is in computer graphics. Nov 28, 2013 at 14:33
• Have you tried googling "manifold learning"?
– user4536
Jan 14, 2014 at 18:19
• I am also interested in the applications of differential geometry on artificial intelligence; however, it is hard, at least to me, to find the applications which has been applied for human beings. I am studying "information geometry" and using the book: Shun'ichi Amari, Hiroshi Nagaoka - Methods of information geometry, Translations of mathematical monographs; v. 191, American Mathematical Society, 2000. I am thinking it is possible to make something interesting. Any ideas are welcome to discuss.
– user126302
Feb 5, 2014 at 3:02

For applications of Differential Geometry in Computer Science, the following link is very useful: http://stat.fsu.edu/~anuj/CVPR_Tutorial/ShortCourse.htm

It talks about : "Differential Geometric Methods for Shape Analysis and Activity Recognition"

One of the contributors of the above is Dr.Anuj Srivastava (http://stat.fsu.edu/~anuj/index.php). I liked the following line from his research areas page- "This technology is needed in many applications -- medical diagnosis, biometrics, video surveillance, undersea imaging, terrain mapping, and satellite image analysis."

Hope it helps !

There are certainly several areas in which differential geometry (DG) affects AI, particularly within machine learning and computer vision.

-- Data as Manifolds --

The basic idea of "manifold learning", which is a technique largely used for non-linear dimensionality reduction (NLDR), is to view the data as lying on a low-dimensional (Riemannian) manifold in a high-dimensional raw data space. Now, there are many methods for NLDR that are not really differential geometric in nature but are still called manifold learning methods. However, there are a few very "DG" methods.

One is Laplacian Eigenmaps. Every Riemannian manifold has an associated Laplace-Beltrami operator, with an accompanying operator spectrum. This spectrum has very good properties for representing functions on the manifold (see: Belkin & Niyogi's 2003 paper, and also "On the optimality of shape and data representation in the spectral domain" by Aflalo, Brezis, and Kimmel, 2015), and are thus a good choice for use as an embedding representation function. The Laplace-Beltrami operator is estimated via a discrete nearest neighbours graph Laplacian that converges to the true continuous operator of the underlying manifold, from which the data is sampled, in the limit.

Another one is Diffusion Maps. Here, one looks at the distance between points using an approximation of the "diffusion distance" on the underlying manifold. This is of course closely linked to the diffusion (heat) PDE on the manifold.

But, does data actually lie on low dimensional manifolds? There is some work on testing this, such as Fefferman et al, 2016, Testing the Manifold Hypothesis.

-- Information Geometry of Statistical Models --

If you consider a family of parameterized probability distributions, then an obvious way to measure distance between them is to look at e.g. the distance between the parameter values. However, this is actually a terrible way to measure this distance. Instead, one should look at how, say, the KL divergence between the distributions change, as one varies the parameters away from each other. This is (roughly) the idea of information geometry. See also here.

One essentially considers the space of probability distributions formed by some parametrized family, and then defines a Riemannian metric on the space (causing it to become a Riemannian manifold).

Of course, many statistical models are essentially probability distributions. In machine learning and pattern recognition, the goal is to compute the parameters. BUT, now that the parameter space is actually a Riemannian manifold, instead of following the "classic" gradient (which only considers Euclidean distance between parameters), you can follow the "natural gradient", which has vastly superior properties. See Amari, Natural gradient works efficiently in learning. It's main disadvantage is computation time, but people are looking at ways to fix that (e.g. see the work by Martens and Grosse). See also the work by Ollivier on Riemannian Metrics for Neural Networks.

The founder of the field, Amari, also discusses applications to ML in his book Information Geometry and Its Applications.

-- Shape Understanding --

Particularly in 3D computer vision and in efforts to apply machine learning to computer graphics, differential geometry plays a key role. The entire field of Geometric Deep Learning hinges on it. The issue arises because the data itself is manifold-valued, i.e. standard convolutional networks don't really apply to non-Euclidean data. A great reference is Geometric deep learning: going beyond Euclidean data by Bronstein et al.

-- Curvature of Generative Models --

Recent deep generative models learn mappings between a data space and a latent space. In both cases, the surfaces formed within each space can be viewed as Riemannian manifolds. This suggests geometric ways to analyze or improve the algorithms. For example:

• Shao et al, The Riemannian Geometry of Deep Generative Models

• Arvanitidis et al, Latent Space Oddity: On the Curvature of Deep Generative Models