riemannian geometry and machine learning for non
play

Riemannian Geometry and Machine Learning for Non Euclidean Data - PowerPoint PPT Presentation

Riemannian Geometry and Machine Learning for Non Euclidean Data Frank C. Park and C.J. Jang Seoul National University Carl Friedrich Gauss (1777 1855) 15 th Century Mapmaking ...were shortest paths on the sphere (but in most cases theyre


  1. Riemannian Geometry and Machine Learning for Non ‐ Euclidean Data Frank C. Park and C.J. Jang Seoul National University

  2. Carl Friedrich Gauss (1777 ‐ 1855)

  3. 15 th Century Mapmaking

  4. ...were shortest paths on the sphere (but in most cases they’re not) It would be nice if straight lines on maps...

  5. Google Maps (Mercator projection)

  6. Mercator maps are very accurate for countries near the equator (e.g., Brazil)

  7. Greenland vs Africa: Sizes on Mercator Map

  8. Greenland vs Africa: Actual Size Comparison

  9. Mercator Map

  10. Gall ‐ Peters Map

  11. Gall ‐ Peters Map: Greenland vs Africa Relative areas are accurate, but shapes are now distorted

  12. National Geographic Map (Winkel map)

  13. David Hilbert (1862 ‐ 1943)

  14. A Hierarchy of Mappings  Isometry (distortion ‐ free)  Area ‐ preserving  Geodesic ‐ preserving  Angle ‐ preserving (conformal)  ....

  15. Calculus on the Sphere The unit two ‐ sphere is parametrized as � � � � � � � � � 1. Spherical coordinates : x � cos � sin � y � sin � sin � � � cos � Other coordinate parametrizations are possible, e.g., stereographic projection: � � �1 � � � � � � 2� 2� � � 1 � � � � � � , � � 1 � � � � � � , 1 � � � � � �

  16. Calculus on the Sphere �� � , � � , � � � Given a curve on the sphere, its incremental arclength is �� � � �� � � �� � � �� � � �� � � sin � � �� � sin � � �� 0 � �� �� �� 0 1 The matrix � � sin � � 0 1 is called the first fundamental 0 form in classical differential geometry (we’ll call it the Riemannian metric ).

  17. Calculus on the Sphere Calculating lengths and areas on the sphere using spherical coordinates: � � ��  Length of � �� � � �� � sin � � � �� � � � �� � � sin � �� ��  Area of � � Note that the area element is

  18. Calculus on the Sphere: The Setup So Far  Local coordinates :  The Riemannian metric : ���, �� � sin � � 0 0 1  Note 1: Other local coordinates are possible.  Note 2: Other choices of Riemannian metric are also � differently, e.g., choose any possible by defining symmetric positive ‐ definite 3x3 matrix � �� ��, �, �� and set � �� � �� � �� �� �� � � �� �� �� � �� � �� � �� �� � �� � �� � �� ��

  19. Calculus on Riemannian Manifolds A differentiable manifold is a space that is locally diffeomorphic* to Euclidean space (e.g., a multidimensional surface) Manifold � local coordinates x *Invertible with a differentiable inverse. Essentially, one can be smoothly deformed into the other.

  20. Calculus on Riemannian Manifolds A Riemannian metric is an inner product defined on each tangent space that varies smoothly over . �� � � � � � �� � �� � �� � � � � �� � � � �� � � ∈ � ��� symmetric positive-definite

  21. Calculus on Riemannian Manifolds  Length of a curve on (local coordinates � : � � � � �  Volume of a subset of : Volume � � � �

  22. Mappings Between Riemannian Manifolds �� � � � � � �� � �� � �� � �� � � � � � �� � �� � �� � � � � �

  23. Isometry Given two manifolds � and � , the mapping �: � → � is an isometry if it preserves distances and angles everywhere: ���� � � � , � � � ���� � �� � � , ��� � �� , for all � � , � � in � � and � are then said to be isometric to each other; � can be transformed into � without any stretching or tearing. � isometric to � � not isometric to � Original �

  24. Isometry: Mathematical Formulation ���� �: � → � � Coordinates � � Coordinates � ���� metric ���� � metric ���� � � � �� � � � � � � � � ���� � � ������ �� �� � � � �� � � � � �� � � Isometry ⟺ � � � � ���� �� ���� , � � � � � ���� � � � � �� ∈ � ���

  25. Isometries and Gaussian Curvature � There is no isometry � � between manifolds � � of different Gaussian curvatures. What’s � � the best one can do in this case? � ��� � � � � ��� � � ��� � �

  26. Finding Nearly Isometric Maps ���� �: � → � � � ���� � � Local coordinates �, metric ���� Local coordinates �, metric ���� � � � Note : The “distance” must be coordinate ‐ invariant .

  27. Coordinate ‐ Invariance This is Spinal Tap (1984)

  28. Coordinate ‐ Invariant Functionals ��, �� , local coord. � � �� � , ⋯ , � � � Riemannian metric � � �� �� � ��, �� , local coord. � � �� � , ⋯ , � � � Riemannian metric � � �� �� � A coordinate ‐ invariant functional of �: � → � has the general form �� � ⋯ �� � � � � � ��� � , ⋯ , � � � det � � where ��·� is any symmetric function, and � � , ⋯ , � � are the roots of �� � ��� . � �� �

  29. Harmonic Maps Intuition : Take � to be made of elastic (e.g., rubber) and � to be  rigid (e.g., made of steel). Wrap the elastic � so that it covers  � all of �, and and let � settle to its � elastic equilibrium state. This is the harmonic map solution [Eells and � Sampson 1964].

  30. Harmonic Maps: Formulation �  � � � , ⋯ , � � � ∑ � � , with boundary conditions �� � ����� ���  The harmonic mapping functional is det ���� �� � ⋯ �� � � � � � �� � � � ��� � ��������� �� �  Variational equations: � � � � �� � � �� � �� � 1 � �� � � �� det � � � � � �� Γ � � � 0 �� �� � �� � �� � det � ��� ��� ��� ��� � are the Christoffel symbols of the where � �� is ��, �� entry of � �� , Γ �� second kind

  31. Examples of Harmonic Maps Finding the minimum distortion map from the unit interval [0,1] to itself: • Find the mapping that maps the interval � � [0,1] onto [0,1] so as to minimize � • Variational equations are , which correspond to the equations for the line .

  32. Examples of Harmonic Maps Geodesics: Given two points on the Riemannian manifold , find the path of shortest distance connecting these two points:  Find the mapping with endpoints specified � � that minimizes � � � � � � �� � �� � � �  Variational equations: ��� ��� �� �� � �� �� � 0 � � � 1 � � �

  33. Examples of Harmonic Maps Harmonic Functions: Find the equilibrium temperature distribution over a planar region with the boundary temperatures specified: �  Find the mapping with values for specified on the boundary of the region. �  Variational equations: (Laplace’s equation)

  34. Manifold Learning Revisited

  35. Manifold Learning • Find a lower ‐ dimensional, minimum distortion, Euclidean representation of high ‐ dimensional data: � � � � ∈ � � � ��� � � ∈ � � usually � � , � ≪ � • Examples from locally linear embedding (LLE) (Roweis et al. 2000) Face images mapped into 2 ‐ dim space Mapping 3 ‐ dim data to 2 ‐ dim space

  36. Riemannian Manifold Learning • Recall the general setup of our global distortion measure: �� � ⋯ �� � ��� � , ⋯ , � � � det � � � �

  37. Riemannian Manifold Learning Choices need to be made: A classification scheme  Manifolds for existing manifold and learning algorithms  Metric in  Metric H in A roadmap for finding  Integrand function � � new manifold learning  Constraints, boundary conditions methods and  Discretization method algorithms (for example, the harmonic mapping * �� �� � � can be estimated using � � , … , � � , � � � distortion) ��� � � from Laplace ‐ Beltrami operator based method

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend