0.25
0.5
0.75
1.25
1.5
1.75
2
Group Theory and Machine Learning
Published on Mar 03, 200835520 Views
**Machine Learning Tutorial Lecture** The use of algebraic methods—specifically group theory, representation theory, and even some concepts from algebraic geometry—is an emerging new direction in mach
Related categories
Chapter list
Group theoretical methods in Machine Learning00:00
Motive02:03
(G, · ) is a group if:02:36
Symmetry04:43
g : X -> X05:24
g : X -> X06:13
For any g 1, g 2 E G , g 2 g 1 E G .07:16
For any g 1 , g 2 , g 3 E G , g 3 ( g 2 g 1 ) = ( g 3 g 2 ) g 1 .07:42
There exists an identity e E G such that eg = ge = g for any g E G .08:31
For any g E G there is a g-1 E G such that g −1g = e .08:45
A set G endowed with multiplication is a group if09:15
Example 111:28
Example 212:04
Example 312:59
Example 413:53
Example 514:05
Example 615:29
Example 716:07
Example 816:27
Example 917:29
Example 1018:37
Example 1118:47
Example 1219:15
Example 1319:59
Finite groups, Infinite groups21:24
Groups22:56
Representations23:45
Example25:55
The idea is to “model” groups by...31:16
Example31:31
What are the “fundamental” representations?31:49
Two representations are equivalent if32:13
Maschke/Wedderburn theorem35:00
The Clebsch-Gordan decomposition35:49
Harmonic analysis37:38
The Fourier transform on a group is38:17
Classical properties41:01
How about right translation f ( z ) ( x ) = f ( x z ? −1)42:35
f(z)(p) = f(p) p(z) so right-translation preserves the subspaces spanned by the rows of f(p).47:27
The Fourier transform F : f -> f is an isomorphism50:23
Application:51:28
The problem: translation and rotation invariance (1)53:11
The problem: translation and rotation invariance (2)55:39
The classical bispectrum58:06
The problem58:39
Start with Fourier Transform59:32
The Power Spectrum59:41
Note translation property01:02:15
Problem: We lose an awful lot of information01:02:20
The Bispectrum01:02:21
It is the Fourier transform of the triple correlation01:03:32
Applications of the classical bispectrum:01:03:34
Want to do the same for I S O +(2) ... 01:04:58
Non-commutative bispectra01:05:00
Under translations01:05:03
Therefore, the spectrum01:05:41
Now try and couple different components01:08:28
In general, p1(x) ! p2(x) decomposes in the form01:11:10
The bispectrum on a compact group is defined01:12:03
Kakarala01:13:36
X is a homogeneous space of G if G acts on x and x = g(x0) sweeps out the whole of X - Example (1)01:20:33
X is a homogeneous space of G if G acts on x and x = g(x0) sweeps out the whole of X - Example (2)01:21:46
X is a homogeneous space of G if G acts on x and x = g(x0) sweeps out the whole of X - Example (3)01:21:48
X is a homogeneous space of G if G acts on x and x = g(x0) sweeps out the whole of X - Example (4)01:21:49
X is a homogeneous space of G if G acts on x and x = g(x0) sweeps out the whole of X - Example (5)01:21:53
Bispectral invariants for ISO+(2)01:21:55
We have:01:22:09
We want: (1)01:22:22
We want: (2)01:23:11
We want: (3)01:23:36
We want: (4)01:23:48
We want: (5)01:23:51
... locally the action of ISO+(2) on R2 is isomorpic to the action of SO(3) on S2.01:24:31
... the algorithm01:24:33
The Projection01:24:35
The Fourier transform (1)01:25:45
The Fourier transform (2)01:26:07
The bispectrum01:26:15
Results (1)01:29:58
Results (2)01:30:00
Results (3)01:33:08