Target audience: Advanced
Estimated reading time: 5'
Lie groups play a crucial role in Geometric Deep Learning by modeling symmetries such as rotation, translation, and scaling. This enables non-linear models to generalize effectively for tasks like object detection and transformations in generative models.
Table of contents
What you will learn: How to implement and evaluate key operations on 3-dimension Special Orthogonal Lie Group.
Notes:
- This post is a follow up on articles related to differential geometry and geometry learning [ref 1, 2, 3 and 4]
- Environments: Python 3.11, Matplotlib 3.9, Geomstats 2.8.0
- Source code is available at Github.com/patnicolas/Data_Exploration/Lie
- To enhance the readability of the algorithm implementations, we have omitted non-essential code elements like error checking, comments, exceptions, validation of class and method arguments, scoping qualifiers, and import statement.
Disclaimer : A thorough tutorial and explanation of Lie groups, Lie algebras, and geometric priors for deep learning models is beyond the scope of this article. Instead, the following sections concentrate on experiments involving key elements and operations on Lie groups using the Geomstats Python library [ref 5].
Overview
Let's start with some basic definitions
Lie manifolds
A smooth manifold is a topological space that locally resembles Euclidean space and allows for smooth (infinitely differentiable) transitions between local coordinate systems. This structure allows for the use of calculus on the manifold.
The tangent space at a point on a manifold is the set of tangent vectors at that point, like a line tangent to a circle or a plane tangent to a surface.
Tangent vectors can act as directional derivatives, where you can apply specific formulas to characterize these derivatives.
In differential geometry, a Lie group is a mathematical structure that combines the properties of both a group and a smooth manifold. It allows for the application of both algebraic and geometric techniques. As a group, it has an operation (like multiplication) that satisfies certain axioms (closure, associativity, identity, and invertibility) [ref 6].
A 'real' Lie group is a set G with two structures: G is a group and G is a (smooth, real) manifold. These structures agree in the following sense: multiplication (a.k.a. product or composition) and inversion are smooth maps.
A morphism of Lie groups is a smooth map which also preserves the group operation: f(gh) = f(g)f(h) and f(1) = 1.
A morphism of Lie groups is a smooth map which also preserves the group operation: f(gh) = f(g)f(h) and f(1) = 1.
Fig. 2 Manifold with tangent space and identity and group element
(Courtesy A. Kirillov Jr Department of Mathematics SUNY at Stony Brook)
Why Lie groups
Lie groups have numerous practical applications in various fields:
- Physics: They describe symmetries in classical mechanics, quantum mechanics, and relativity.
- Robotics: Lie groups model the motion of robots, particularly in the context of rotation and translation (using groups like SO(3) and SE(3)).
- Control Theory: Lie groups are used in the analysis and design of control systems, especially in systems with rotational or symmetrical behavior.
- Computer Vision: They help in image processing and 3D vision, especially in tasks involving rotations and transformations.
- Differential Equations: Lie groups are instrumental in solving differential equations by leveraging symmetry properties.
In term of machine learning...
- Geometric Deep Learning: Lie groups help capture rotational, translational, or scaling symmetries, making models more efficient and generalizable in tasks like image recognition and 3D object detection.
- Generative Models: Lie groups allow for more structured latent spaces, enabling better control over transformations in generative models like GANs and VAEs.
- Reinforcement Learning: They are used to model continuous action spaces, improving control over robotic systems.
- Optimization: Lie groups help design efficient optimization techniques on curved spaces, like Riemannian manifolds.
Examples of Lie groups
Group of Invertible 2 x 2 matrices of real values
Group of Invertible 2 x 2 matrices of complex values
Group of Invertible 3 x 3 matrices of real values
Special Unitary group 2 x 2 matrices with determinant 1
Special Unitary group 3 x 3 matrices with determinant 1
Special Orthogonal (3D rotation, 2x2 matrices) group
Special Orthogonal (3D rotation, 3x3 matrices) group
Special Euclidean group
Special Orthogonal Group
The Special Orthogonal Group in 3 dimensions, SO(3) is the group of all rotation matrices in 3 spatial dimensions.
It can be defined by 3 rotation elements for each of the axis of rotation x, y, and z.
Implementation
Geomstats is a free, open-source Python library designed for conducting machine learning on data situated on nonlinear manifolds, an area known as Geometric Learning. This library offers object-oriented, thoroughly unit-tested features for fundamental manifolds, operations, and learning algorithms, compatible with various execution environments, including NumPy, PyTorch, and TensorFlow (Overview Geomstats library).
The library is structured into two principal components:
- geometry: This part provides an object-oriented framework for crucial concepts in differential geometry, such as exponential and logarithm maps, parallel transport, tangent vectors, geodesics, and Riemannian metrics.
- learning: This section includes statistics and machine learning algorithms tailored for manifold data, building upon the scikit-learn framework.
First let's wrap the element of a point or matrix into a class, SO3Point with the following attributes:
- group_element Point on the manifold
- base_point on the manifold (Identity if undefined)
- description which describes the rotation matrix
from dataclasses import dataclass
@dataclass
class SO3Point:
group_element: np.array
base_point: np.array # Default identity matrix
descriptor: AnyStr
Secondly, let's build a class ,LieSO3Group, that encapsulates the definition of the Special orthogonal group of dimension 3 and its related operations.
We specify two constructors:
- __init__: Default constructor that create a new element in the SO3 manifold, group_element, using a tangent vector (3x3 rotation matrix) tgt_vector and identity, base_point.
- build: Alternative constructor for which the tangent vector is a 9- element list and a tuple as base point.
import geomstats.backend as gs
from geomstats.geometry.special_orthogonal import SpecialOrthogonal
class LieSO3Group(object):
dim = 3
# Lie group as defined in Geomstats library
lie_group = SpecialOrthogonal(n=dim, point_type='vector', equip=False)
identity = gs.eye(dim) # Define identity
def __init__(self, tgt_vector: np.array, base_point: np.array = identity) -> None
self.tangent_vec = gs.array(tgt_vector)
# Exp. a left-invariant vector field from a base point
self.group_element = LieSO3Group.lie_group.exp(self.tangent_vec, base_point)
self.base_point = base_point
@classmethod
def build(cls, tgt_vector: List[float], base_point: List[float] = None) -> Self:
np_input = np.reshape(tgt_vector, (3, 3))
np_point = np.reshape(base_point, (3, 3)) if base_point is not None
else LieSO3Group.identity
return cls(tgt_vector=np_input, base_point=np_point)
We use the constructors of the LieSO3Group class to generate two points on the SO3 manifold, so3_point1 and so3_point2. These points are represented as vectors in 3-dimensional Euclidean space for visualization purposes.
The first point use the identity as base point for the tangent space of rotation of 90 degrees around X axis:
[[ 1. 0. 0.]
[ 0. 0. -1.]
[ 0. 1. 0.]]
The second SO3 point used the same tangent vector with a base
[[ 0. -1. 0.]
[ 1. 0. 0.]
[ 0. 0. 1.]]
so3_tangent_vec = [1.0, 0.0, 0.0, 0.0, 0.0, -1.0, 0.0, 1.0, 0.0]
# First rotation matrix +90 degrees around X-axis with identity as base point
so3_group = LieSO3Group.build(so3_tangent_vec)
so3_point1 = SO3Point(
group_element=so3_group.group_element,
base_point=LieSO3Group.identity,
descriptor='SO3 point from tangent vector\n[1 0 0]\n[0 0 -1]\n[0 1 0]\nBase point: Identity')
# Same rotation matrix with another rotation matrix 90 degrees around
# Y-axis for base point base_point = [0.0, -1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0] so3_group2 = LieSO3Group.build(so3_tangent_vec, base_point) so3_point2 = SO3Point( group_element=so3_group2.group_element, base_point=base_point, descriptor='Same SO3 point\nBase point:\n[0 -1 0]\n[0 0 0]\n[0 0 1]')
LieSO3Group.visualize_all([so3_point1, so3_point2])
so3_point1:
[[ 2.00 0.00 0.00] [[2 0 0]
[ 0.00 0.93 -0.99] approximation of [0 1 -1]
[ 0.00 0.99 0.93]] [0 1 1]]
so3_point2:
[[ 0.99 -0.93 0.00] [[1 -1 0]
[ 0.93 0.00 -0.99] approximation of [1 0 -1]
[ 0.00 0.99 0.93]] [0 1 1]
The code for the visualization methods are available on Github at LieSO3Group
Inverse rotation
The first operation to illustrate is the inverse rotation using Geomstats library. It is implemented by the LieSO3Group method inverse.
def inverse(self) -> Self:
inverse_group_element = LieSO3Group.lie_group.inverse(self.group_element)
return LieSO3Group(inverse_group_element)
For simplicity sake, we create a SO3 group element at identity and visualize its inverse on the same base point.
# Original rotation matrix +90 degrees around X-axis with identity as base point
\so3_tangent_vec = [1.0 0.0,0.0, 0.0, 0.0, -1.0, 0.0, 1.0, 0.0]
so3_group = LieSO3Group.build(so3_tangent_vec)
so3_point = SO3Point(
group_element=so3_group.group_element,
base_point=LieSO3Group.identity,
descriptor='SO3 point from tangent vector\n[1 0 0]\n[0 0 -1]\n[0 1 0]\nBase point: Identity')
# Inverse SO3 rotation matrix
so3_inv_group = so3_group.inverse()
inv_so3_point = SO3Point( group_element=so3_inv_group.group_element, base_point=LieSO3Group.identity,
descriptor='SO3 inverse point')
As expected, the inverse of an SO3 group element represents the inverse rotation in 3-dimensional space. This would not hold true if the base point were not the identity element.
Original SO3 point
[[ 2.00 0.00 0.00]
[ 0.00 0.93 -0.99]
[ 0.00 0.99 0.93]]
SO3 Inverse point:
[[-1.00 0.00 0.00]
[ 0.00 0.00 0.93]
[ 0.00 -0.93 0.00]]
Fig. 4 Visualization of the inverse of SO3 element at identity
Composition
As previously mentioned, a smooth manifold with the structure of a Lie group ensures that the product (or composition) of two elements also belongs to the manifold. The following LieSO3Group method, product, utilizes the `compose method of Geomstats library.
def product(self, lie_so3_group: Self) -> Self:
composed_group_point = LieSO3Group.lie_group.compose(self.group_element lie_so3_group.group_element)
return LieSO3Group(composed_group_element)
In this initial test, we use the rotation matrix of 90 degrees around Z-axes as the second component for the composition.
# Second SO3 rotation matrix +90 degree along Z-axis so3_tangent_vec2 = [0.0, -1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0]
so3_group2 = LieSO3Group.build(tgt_vector=so3_tangent_vec2)
so3_point2 = SO3Point(
group_element=so3_group2.group_element,
base_point=LieSO3Group.identity
descriptor='SO3 point from tangent vector\n[0 -1 0]\n[1 0 0]\n[0 0 1]\nBase point: Identity')
# Composition of two rotation matrices SO3 group
so3_group_product = so3_group.product(so3_group2)
Composition of rotation matrix +90 degrees along X-axis with rotation matrix +90 degrees along Z-axis:
[[-2.11 0.62 0.96]
[ 1.52 1.74 -1.52 ]
[-0.96 -0.62 -2.11 ]]
As expected, the composition produces an anti-symmetric matrix
Fig. 5 Visualization of the composition of a SO3 group element with its inverse at identity
The second evaluation consists of composing the initial SO3 group element (90 degrees rotation along X axis) with identity matrix.
[[-1.28 0.00 0.00 ]
[-0.96 -2.11 0.62]
[-0.96 -0.62 -2.11]]
Fig. 6 Visualization of the composition of two SO3 group elements at identity
References
----------------------------------
Patrick Nicolas has over 25 years of experience in software and data engineering, architecture design and end-to-end deployment and support with extensive knowledge in machine learning.
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning", Packt Publishing ISBN 978-1-78712-238-3 and Geometric Learning in Python Newsletter on LinkedIn.
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning", Packt Publishing ISBN 978-1-78712-238-3 and Geometric Learning in Python Newsletter on LinkedIn.