forked from akmaier/pr-slides
-
Notifications
You must be signed in to change notification settings - Fork 0
/
19_ica.tex
executable file
·1216 lines (947 loc) · 37.1 KB
/
19_ica.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
\def\kurt{\mathop{\mathsf{kurt}}}
\def\MI{\mathop{\mathsf{MI}}}
\def\KL{\mathop{\mathsf{D}_{\mathsf{KL}}}}
\section{Independent Component Analysis}
\subsection{Motivation}
\begin{frame}
\frametitle{Cocktail-Party Problem}
\begin{ovalblock}{Example}
Imagine the following situation:
\begin{itemize}
\item You have two microphones in a room at different locations.
\item The microphones record time signals $x_1(t), x_2(t)$. \pause
\item Each recorded signal is a weighted sum of two speakers $s_1(t), s_2(t)$: \pause
\end{itemize}
%
\begin{eqnarray*}
x_1(t) &=& a_{11} s_1(t) + a_{12} s_2(t) \\
x_2(t) &=& a_{21} s_1(t) + a_{22} s_2(t)
\end{eqnarray*}
Parameters $a_{ij}$ depend on the distance of the microphones to the speakers.
\end{ovalblock}
\end{frame}
\begin{frame}
\frametitle{Cocktail-Party Problem}
\begin{ovalblock}{Example}
For simplicity, we assume just a very simple mixing model without any time delays or other factors. \\[.5cm]
\pause
\structure{Observations:}
\begin{itemize}
\item If we knew the $a_{ij}$, the problem of reconstructing $s_i$ is to solve the linear equations by classical methods. \pause
\item \structure{But:} We do not know the $a_{ij}$! Thus, the problem is considerably more difficult!
\end{itemize}
\end{ovalblock}
\end{frame}
\mode<presentation>{
\begin{frame}
\frametitle{Cocktail-Party Problem}
\begin{ovalblock}{Example}
\footnotesize
Original sound sources:
\vspace*{-0.25cm}
\begin{figure}
\centering
\movie[]{\includegraphics[width=1cm]{\jpgdir/maier.\jpg}}{./audio/sourceX.wav} \qquad
\movie[]{\includegraphics[width=1cm]{\jpgdir/piano.\jpg}}{./audio/sourceY.wav} \qquad
\end{figure}
\pause
%
Samples at the cocktail-party:
\vspace*{-0.25cm}
\begin{figure}
\centering
\movie[]{\includegraphics[width=1cm]{\jpgdir/microphone.\jpg}}{./audio/mixedX.wav} \qquad
\movie[]{\includegraphics[width=1cm]{\jpgdir/microphone.\jpg}}{./audio/mixedY.wav} \qquad
\end{figure}
\pause
%
Reconstructed sound sources:
\vspace*{-0.25cm}
\begin{figure}
\centering
\movie[]{\includegraphics[width=1cm]{\jpgdir/speaker.\jpg}}{./audio/separateX.wav} \qquad
\movie[]{\includegraphics[width=1cm]{\jpgdir/speaker.\jpg}}{./audio/separateY.wav} \qquad
\end{figure}
\end{ovalblock}
\end{frame}
}
\begin{frame}
\frametitle{Cocktail-Party Problem}
The principle for solving the cocktail-party problem has a lot of \\
\structure{other interesting applications}: \\[.25cm]
\begin{itemize}
\item speech signal recovery: telecommunications \\[.25cm]
\item recovery of images from mixed signals: MRI, fMRI \\[.25cm]
\item electrical recordings of brain activity:
\begin{itemize}
\item EEG (electroencephalography)
\item MEG (magnetoencephalography) \\[.25cm]
\end{itemize}
\item feature extraction \\[.25cm]
\item multispectral image analysis
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Separate Natural Images\cont}
\begin{figure}
\copyrightbox[b]{
\makebox[.6\linewidth]{
\includegraphics[width=0.70\linewidth]{\pngdir/fig6.\png}
}
}{Isomura, T., Toyoizumi, T. \href{https://doi.org/10.1038/srep28073}{A Local Learning Rule for Independent Component Analysis. Sci Rep 6, 28073 (2016)}}
%\caption{Final reconstructed images.}
\end{figure}
\end{frame}
\begin{frame}
\frametitle{MEG: Recovery of Brain Activity}
\begin{figure}
%\includegraphics[width=0.6\linewidth]{\pngdir/ica_meg.\png}
\copyrightbox[b]{
\makebox[.6\linewidth]{
\includegraphics[width=0.55\linewidth]{\pngdir/ica_meg.\png}
}
}{R. Vigário, V. Jousmäki, M. Hämäläinen, R. Hari, E. Oja. \href{https://papers.nips.cc/paper/1466-independent-component-analysis-for-identification-of-artifacts-in-magnetoencephalographic-recordings.pdf}{``Independent component analysis for identification of artifacts in magnetoencephalographic recordings, Advances in Neural Information Processing Systems''}}
\caption{Principle of MEG acquisition.}
\end{figure}
\end{frame}
\begin{frame}
\frametitle{MEG: Recovery of Brain Activity \cont}
\begin{figure}
%\includegraphics[height=0.8\textheight]{\pngdir/ica_meg_est.\png}
\copyrightbox[b]{
\makebox[.6\linewidth]{
\includegraphics[height=0.7\textheight]{\pngdir/ica_meg_est.\png}
}
}{R. Vigário, V. Jousmäki, M. Hämäläinen, R. Hari, E. Oja. \href{https://papers.nips.cc/paper/1466-independent-component-analysis-for-identification-of-artifacts-in-magnetoencephalographic-recordings.pdf}{``Independent component analysis for identification of artifacts in magnetoencephalographic recordings, Advances in Neural Information Processing Systems''}}
\caption{Recovered signals.}
\end{figure}
\end{frame}
\begin{frame}
\frametitle{Common Framework}
\structure{Idea:}
\begin{itemize}
\item Use some information about the statistical properties of the signals $s_i(t)$ to estimate $a_{ij}$.
\end{itemize}
\pspread
Surprisingly, it turns out that the only statistical assumption that we have to make is that the $s_i(t)$ are \textit{statistically independent} at each time point $t$.
\pspread
Formulation in a unified mathematical framework (H\'{e}rault and Jutten, 1984-1991):
\begin{center}
\structure{ICA -- Independent Component Analysis}
\end{center}
\end{frame}
\subsection{Latent Variables and Factor Analysis}
\begin{frame}
\frametitle{Latent Variables and Factor Analysis}
\structure{Statistical latent variables model:}
\begin{itemize}
\item Rewrite the time series into $n$ linear mixture observations $x_1, \ldots, x_n$ \pause
\item Each mixture $x_i$ as well as each component $s_j$ are random variables
\end{itemize}
\begin{displaymath}
x_i = \sum_{j=1}^m a_{ij} s_j ~, \qquad i = 1,\ldots,n
\end{displaymath}
\pause
In matrix notation:
\begin{displaymath}
\vec x = \mat A \vec s
\end{displaymath}
where
\begin{itemize}
\item $\mat A$ is a constant \textit{mixing} matrix
\item $s_j$ are latent random variables (independent components)
\item both $\mat A$ and $s_j$ have to be estimated based on observations $x_i$
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{First Approach: Decorrelation}
Assuming $\vec{\bar{x}} = \vec{0}$, we already know a latent variable representation \\
(see chapter \emph{Discriminant Analysis I}). \\[.5cm] \pause
From:
\begin{displaymath}
E\{ \vec x \vec x^T \} = \mat \Sigma = \mat U \mat D \mat U^T
\end{displaymath}
\pause
and
\begin{displaymath}
\mat{\Sigma}^{-1}
= \mat{U}\mat{D}^{-1}\mat{U}^T
= (\mat{U}\mat{D}^{-\frac{1}{2}})\cdot \mat{I} \cdot(\mat{U}\mat{D}^{-\frac{1}{2}})^T
\end{displaymath}
\pause
we compute a mapping:
\begin{displaymath}
\vec{\tilde{x}} = \mat D^{-\frac{1}{2}} \mat U^T \vec x
\end{displaymath}
\end{frame}
\begin{frame}
\frametitle{First Approach: Decorrelation}
For zero-mean vectors, the mapping:
\begin{displaymath}
\vec{\tilde{x}} = \mat D^{-\frac{1}{2}} \mat U^T \vec x
\end{displaymath}
is called \structure{\textit{Whitening Transform}}. \\[0.3cm] \pause
It has some \structure{interesting properties}:
\begin{itemize}
\item The mapped random variables $\tilde{x}_i$ are uncorrelated. \pause
\item $\vec{\tilde{x}}$ has unit variance:
\end{itemize}
\begin{eqnarray*}
E\{ \vec{\tilde{x}} \vec{\tilde{x}}^T \}
&=& E\left\{ \Big( \mat D^{-\frac{1}{2}} \mat U^T \vec x \Big) \Big( \mat D^{-\frac{1}{2}} \mat U^T \vec x \Big)^T \right\} \pause
= E\left\{ \mat D^{-\frac{1}{2}} \mat U^T \vec x \vec x^T \mat U \mat D^{-\frac{1}{2}} \right\} \\ \pause
&=& \mat D^{-\frac{1}{2}} \mat U^T \mat \Sigma \mat U \mat D^{-\frac{1}{2}} = \mat D^{-\frac{1}{2}} \mat U^T \cdot \mat U \mat D \mat U^T \cdot \mat U \mat D^{-\frac{1}{2}} \\ \pause
&=& \mat D^{-\frac{1}{2}} \mat D \mat D^{-\frac{1}{2}} \pause
= \mat I
\end{eqnarray*}
\end{frame}
\begin{frame}
\frametitle{First Approach: Decorrelation}
We could interpret the mapped random variable $\vec{\tilde{x}}$ as an estimate of the latent variable model:
\begin{displaymath}
\vec s = \vec{\tilde x}
\end{displaymath}
But this would give poor results. \\[.3cm] \pause
\structure{Problem:} The whitening transform is not unique! \\[.3cm]
Consider for example an arbitrary orthogonal matrix $\mat R \in \real^{n \times n}$:
%
\small
\begin{eqnarray*}
\vec{\hat x}
&=& \mat R \vec{\tilde x} = \mat R \mat D^{-\frac{1}{2}} \mat U^T \vec x \\[.2cm] \pause
E\{ \vec{\hat x} \vec{\hat x}^T \}
&=& E\left\{ \mat R \left(\mat D^{-\frac{1}{2}} \mat U^T \vec x \right) \left( \mat D^{-\frac{1}{2}} \mat U^T \vec x \right)^T \mat R^T \right\} \\ \pause
&=& \mat R \mat I \mat R^T \\ \pause
&=& \mat I
\end{eqnarray*}
\end{frame}
\input{nextTime.tex}
\begin{frame}
\frametitle{Second Approach: Independence}
\structure{Observations:}
\begin{itemize}
\item Lack of correlation determines the second-degree cross-moments (covariances) of a multi-variate distribution.
\item Statistical independence is stronger, as it determines all of the cross-moments. \\[.3cm] \pause
\end{itemize}
Given 2 statistically independent random variables $y_1, y_2$ and 2 functions $h_1, h_2$:
\footnotesize
\begin{eqnarray*}
E\{ h_1(y_1) \, h_2(y_2) \}
&=& \iint h_1(y_1) \, h_2(y_2) \, p(y_1, y_2) \, \mathsf{d}y_1 \, \mathsf{d}y_2 \\ \pause
&=& \iint h_1(y_1) \, p(y_1) \, h_2(y_2) \, p(y_2) \, \mathsf{d}y_1 \, \mathsf{d}y_2 \\ \pause
&=& \int h_1(y_1) \, p(y_1) \, \mathsf{d}y_1 \int h_2(y_2) \, p(y_2) \, \mathsf{d}y_2 \\ \pause
&=& E\{ h_1(y_1) \} \, E\{ h_2(y_2) \}
\end{eqnarray*}
\end{frame}
\begin{frame}
\frametitle{Second Approach: Independence}
These extra moment conditions allow us to identify the elements of $\mat A$ uniquely. \\[.5cm]
\structure{Case of Gaussian distribution:}
\begin{itemize}
\item Gaussian distribution is determined by its second moments alone.
\item Any Gaussian independent components can be determined only \\
up to a rotation
\end{itemize}
\spread
Therefore, we assume that the $s_i$ are \structure{independent} and \structure{non-Gaussian}.
\end{frame}
\begin{frame}
\frametitle{Whitening Transform}
The whitening transform is usually done before ICA as a pre-processing step: \pause
\begin{itemize}
\item Mixing matrix $\mat A$ has $n^2$ degrees of freedom. \pause
\item Whitening transforms the mixing matrix $\mat A$ into $\mat{\tilde A}$:
\begin{displaymath}
\vec{\tilde x} = \mat D^{-\frac{1}{2}} \mat U^T \mat A \vec s = \mat{\tilde{A}} \vec s \pause
\end{displaymath}
\item The new mixing matrix is \structure{orthogonal}:
\begin{displaymath}
E\{ \vec{\tilde x} \vec{\tilde x}^T \} = \mat{\tilde A} \, E\{\vec s \vec s^T \} \, \mat{\tilde A}^T = \mat I \pause
\end{displaymath}
\item $\mat{\tilde A}$ is orthogonal and has \structure{$n (n-1) / 2$ degrees of freedom}
\end{itemize}
Thus, applying the whitening transform solves roughly half of the problem.
\end{frame}
\subsection{Illustration of ICA}
\begin{frame}
\frametitle{Illustration of ICA}
Consider two independent components with the following uniform distributions:
%
{\footnotesize
\begin{displaymath}
p(s_i) = \begin{cases}
\frac{1}{2\sqrt{3}} \quad &, ~\mbox{if}~|s_i| \leq \sqrt{3} \\
0 \quad &, ~\mbox{otherwise}
\end{cases}
\end{displaymath}
}
\begin{columns}
\pause
\column{.5\linewidth}
\vspace*{-.7cm}
\begin{center}
\resizebox{.9\linewidth}{!}{
\input{\texfigdir/ica1.pstex_t}
}
\end{center}
\column{.5\linewidth}
\pause
\structure{Properties of the joint pdf:}
\begin{itemize}
\item signal components are independent \pause
\item joint pdf is uniform on square \pause
\item zero mean \pause
\item variance is equal to one
\end{itemize}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{Illustration of ICA \cont}
The two independent components are mixed with the matrix:
{\footnotesize
\begin{displaymath}
\mat A = \left(
\begin{array}{cc}
2 & 3 \\
2 & 1
\end{array}
\right) ~\normalsize ,~\mbox{which results in}~ \vec x = \mat A \vec s ~.
\end{displaymath}
}
\begin{columns}
\pause
\column{.5\linewidth}
\vspace*{-.7cm}
\begin{center}
\resizebox{0.9\linewidth}{!}{
\input{\texfigdir/ica2.pstex_t}
}
\end{center}
\column{.5\linewidth}
\pause
\structure{Properties of the mixed signal:}
\begin{itemize}
\item joint pdf of mixed signals is uniform on a parallelogram \\[.3cm] \pause
\end{itemize}
\structure{More important:}
\begin{itemize}
\item $x_1, x_2$ are not independent any more
\end{itemize}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{Illustration of ICA \cont}
\structure{Intuitive way of estimating $\mat A$:}
\begin{itemize}
\item Edges of the parallelogram are in the directions of the columns of $\mat A$. \pause
\item In principle, we could first estimate the joint pdf of $x_1, x_2$. \pause
\item If we locate the edges of the joint pdf, we can estimate $\mat A$.
\end{itemize}
\pspread
\structure{But:}
\begin{itemize}
\item Computationally quite expensive
\item This principle works only with \structure{exactly uniform} distributions.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Illustration of ICA \cont}
\structure{Effect of the whitening transform applied to the data:} \\[.5cm]
\begin{columns}
\column{.5\linewidth}
\vspace*{-0.7cm}
\begin{center}
\resizebox{.9\linewidth}{!}{
\input{\texfigdir/ica_whitened.pstex_t}
}
\end{center}
%
\column{.5\linewidth}
\pause
\structure{Properties of the whitened observations:}
\begin{itemize}
\item Joint pdf of $\vec{\tilde x}$ is uniform on a square. \pause
\item Components are determined except for rotation. \pause
\item Problem of recovering $\mat{\tilde A}$ is much simpler.
\end{itemize}
\end{columns}
\end{frame}
\subsection{Properties of ICA}
\begin{frame}
\frametitle{Basic Properties}
\structure{Assumptions for the ICA model:}
\begin{itemize}
\item We assume that the $s_j$ are mutually independent. \pause
\item The $s_j$ have to be non-Gaussian in order to determine them from the $x_i$. \pause
\item For simplicity, we assume that $\mat A$ is square.
\end{itemize}
\pspread
\structure{Ambiguities of the ICA model:}
\begin{itemize}
\item The $s_j$ are defined only up to a multiplicative constant.
\item The $s_j$ are not ordered.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Ambiguities of ICA}
Writing the ICA model in terms of the colums of $\mat A$:
\begin{displaymath}
\vec x = \sum_{i=1}^n \vec a_i s_i \pause
\end{displaymath}
\begin{itemize}
\item Any \structure{scalar multiplier} for $s_i$ can be eliminated by scaling $\vec a_i$ appropriately. \pause
\item The matrix $\mat A$ can be adapted to restrict the $s_i$ to have \structure{unit variance}. \pause
\item This still leaves the ambiguity of the \structure{sign}: \\
multiplying $s_i$ by $\pm 1$ does not affect the model. \pause
\item This ambiguity is usually \structure{insignificant} in most applications.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Ambiguities of ICA \cont}
\structure{Ambiguity of the ordering:}
\begin{displaymath}
\vec x = \sum_{i=1}^n \vec a_i s_i
\end{displaymath}
\pspread
\begin{itemize}
\item As both $\vec s$ and $\mat A$ are unknown we can change the order of the summation. \pause
\item Formalization using a \structure{permutation matrix $\mat P$}:
\begin{displaymath}
\vec x = \underbrace{\mat A \mat P}_{\mat A^{\ast}} \underbrace{\mat P^{-1} \vec s}_{\vec s^{\ast}} \pause
\end{displaymath}
\item $\mat A^{\ast}$ is just a new mixing matrix to be solved.
\end{itemize}
\end{frame}
\subsection{Basic Principle of ICA}
\begin{frame}
\frametitle{Basic Principle of ICA}
So far, if we know $\mat A$, we could compute its inverse $\mat A^{-1}$ to obtain the independent components. Consider a linear combination of $x_i$ with a weight vector $\vec w$:
\begin{displaymath}
y = \vec w^T \vec x
\end{displaymath}
\spread
Clearly, $y$ equals one of the independent components if $\vec w$ is one row of $\mat A^{-1}$.
\end{frame}
\begin{frame}
\frametitle{Basic Principle of ICA \cont}
Change in variables:
\begin{displaymath}
\vec z = \mat A^T \vec w
\end{displaymath}
Applied to the linear combination:
\begin{displaymath}
y = \vec w^T \vec x = \vec w^T \mat A \vec s = \vec z^T \vec s
\end{displaymath}
\pspread
\structure{Result from the Central Limit Theorem:}
\begin{itemize}
\item The sum of a number of independent random variables tends toward a normal distribution. \pause
\item $\vec z^T \vec s$ is \textit{more Gaussian} than any of the $s_i$ \pause
\item $\vec z^T \vec s$ is \textit{least Gaussian} when it equals one of the $s_i$
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Basic Principle of ICA \cont}
\structure{Key principle of ICA:} \\[.25cm]
Maximizing the non-Gaussianity of $\vec w^T \vec x$ results in the independent components!
\end{frame}
\begin{frame}
\frametitle{Basic Principle of ICA \cont}
\structure{Marginal distributions of the joint and the mixed signal:}
\begin{center}
\resizebox{\linewidth}{!}{%
\only<1>{\input{\texfigdir/ica1.pstex_t}}%
\only<1>{\hspace{0.5cm}\input{\texfigdir/ica1_marginal.pstex_t}}%
\only<2>{\input{\texfigdir/ica2.pstex_t}}%
\only<2>{\hspace{0.5cm}\input{\texfigdir/ica2_marginal.pstex_t}}%
}
\end{center}
\end{frame}
\begin{frame}
\frametitle{Basic Principle of ICA \cont}
\structure{Reasons for the importance of the non-Gaussianity:} \pause
\begin{itemize}
\item In case of Gaussian random variables, the ICA model can only be estimated up to an orthogonal transformation. \\[0.25cm]
\structure{Note:}\\
\quad If just one of the components is Gaussian, ICA still works. \\[0.25cm] \pause
\item The Gaussian is the \textit{most random} distribution within the family of pdfs with given mean and variance. \pause
\item Therefore, it is the least informative pdf with respect to the underlying data.
\end{itemize}
\pspread
Distributions that have the \textit{least resemblance} to the Gaussian reveal more structure associated with the data.
\end{frame}
\begin{frame}
\frametitle{Importance of Non-Gaussianity}
The \textit{randomness} can be measured using the concept of entropy from Shannon's information theory:
\begin{itemize}
\item Entropy is a measure of the uncertainty of an event, or the randomness of a measure.
\end{itemize}
\spread
\begin{citeblock}{Differential Entropy}
The \textit{differential entropy} $H(X)$ of a continuous random variable $X$ with density $p(x)$ is defined as
\begin{displaymath}
H(p) = - \int p(x) \log p(x) \,\mathsf{d}x
\end{displaymath}
\end{citeblock}
% The \textit{Kullback-Leibler distance} or \textit{relative entropy} between two densities $p$ and $q$ is:
% $$
% D(p \| q) = \int p(x) \log \frac{p(x)}{q(x)} \mathsf{d}x
% $$
\end{frame}
\begin{frame}
\frametitle{Importance of Non-Gaussianity \cont}
\begin{citeblock}{Theorem}
The Gaussian maximizes the entropy over all distributions with the same mean and the same covariance.
\end{citeblock}
\pspread
\structure{Proof:} \pause
\begin{itemize}
\item Let $x$ be the random variable, $p(x)$ the pdf that has the highest randomness. \pause
\item Rewrite moments $M_i$ equations using a set of polynomials $\{ r_i(x) \}$:
\begin{displaymath}
\int p(x) r_i(x) \,\mathsf{d}x = M_i ~,~\mbox{where}~M_i~\mbox{are called moments}. \pause
\end{displaymath}
\item Using $r_0(x) \equiv 1$, $M_0 = 1$ constrains $p(x)$ to be a pdf.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Importance of Non-Gaussianity \cont}
\structure{Lagrangian functional for the maximum entropy problem:}
{\small
\begin{displaymath}
\argmin_{p(x)} J \equiv
\argmin_{p(x)} \int p(x) \log p(x) \,\mathsf{d}x - \sum_{i = 0}^N \lambda_i \left( \int p(x) r_i(x) \,\mathsf{d}x - M_i \right)
\end{displaymath}
}
\pause
Taking the functional derivative with respect to $p(x)$ (G\^{a}teaux derivative) and setting it to zero:
%
{\small
\begin{displaymath}
\frac{\delta J}{\delta p} = \log p(x) + 1 - \sum_{i=0}^N \lambda_i r_i(x) \stackrel{!}{=} 0
\end{displaymath}
}
\pause
%
yields the family of exponential distributions:
%
{\small
\begin{displaymath}
p(x) = \exp \left( -1 + \sum_{i=0}^N \lambda_i r_i(x) \right)
\end{displaymath}
}
\end{frame}
\begin{frame}
\frametitle{Importance of Non-Gaussianity \cont}
Result for using first and second moments for mean $\mu$ and variance $\sigma^2$:
\begin{displaymath}
p(x) = e^{-(1 - \lambda_0 - \lambda_1 x - \lambda_2 (x-\mu)^2)}
\end{displaymath}
\pause
Plug the form into the constraints:
\begin{eqnarray*}
\int e^{-(1 - \lambda_0 - \lambda_1 x - \lambda_2 (x-\mu)^2)} \,\mathsf{d}x &=& 1 \\
\int x e^{-(1 - \lambda_0 - \lambda_1 x - \lambda_2 (x-\mu)^2)} \,\mathsf{d}x &=& \mu \\
\int (x-\mu)^2 e^{-(1 - \lambda_0 - \lambda_1 x - \lambda_2 (x-\mu)^2)} \,\mathsf{d}x &=& \sigma^2
\end{eqnarray*}
\end{frame}
\begin{frame}
\frametitle{Importance of Non-Gaussianity \cont}
Integrate analytically (non-trivial) and solve for Lagrangian multipliers:
\begin{eqnarray*}
\lambda_0 &=& 1 - \frac{1}{2} \log(2 \pi \sigma^2 ) \\
\lambda_1 &=& 0 \\
\lambda_2 &=& - \frac{1}{2 \sigma^2}
\end{eqnarray*}
\pspread
Insert the results into the form of $p(x)$:
\begin{eqnarray*}
p(x) &=& e^{-(1 - \lambda_0 - \lambda_1 x - \lambda_2 (x-\mu)^2)} \\
&=& e^{-\frac{1}{2} \log(2 \pi \sigma^2)} e^{-\frac{(x-\mu)^2}{2 \sigma^2}} \\
&=& \frac{1}{\sqrt{2 \pi} \sigma} e^{- \frac{(x-\mu)^2}{2 \sigma^2}}
\end{eqnarray*}
\hfill \qed
\end{frame}
\subsection{ICA Estimation Algorithm}
\begin{frame}
\frametitle{ICA Estimation Algorithm}
\structure{Basic ICA Estimation Algorithm:}
\begin{centernss}
\resizebox{.9\linewidth}{!}{
\begin{struktogramm}(105,50)
\assign{Apply centering transform}
\assign{Apply whitening transform}
\assign{$i \gets 1$}
\until{$i > n$~ ($n$: number of independent components)}
\assign{Take a random vector $\vec w_i$}
\assign{Maximize non-Gaussianity of $\vec w_i^T \vec x$ subject to\\
$\| \vec w_i \| = 1$ \\
$\vec w_j^T \vec w_i = 0 ~,~ j < i$}
\assign{$i \gets i+1$}
\untilend
\assign{Use weight matrix: $\mat W = \left( \vec w_1^T, \vec w_2^T, \ldots, \vec w_n^T \right)$ to compute $\vec s$}
\assign{Output: independent components $\vec s$}
\end{struktogramm}
}
\end{centernss}
\end{frame}
\begin{frame}
\frametitle{ICA Estimation Algorithm \cont}
\structure{Notes:}
\begin{itemize}
\item Estimation by maximizing non-Gaussianity of independent components. \pause
\item There exist equivalent algorithms for solving the ICA:
\begin{itemize}
\item Gradient descent methods
\item Fast ICA
\end{itemize}
\pause
\item Relation to Projection Pursuit approach (Friedman and Tukey, 1974):
\begin{itemize}
\item Projection Pursuit is a method for visualization and exploratory data analysis.
\item Attempts to show clustering structure by finding \textit{interesting} projections.
\item Interestingness is usually measured by non-Gaussianity.
\end{itemize}
\end{itemize}
\end{frame}
\input{nextTime.tex}
\subsection{Measures of Non-Gaussianity}
\begin{frame}
\frametitle{Measures of Non-Gaussianity}
\begin{itemize}
\item So far, we have seen that the key principle in estimating independent components is the non-Gaussianity. \pause
\item In order to optimize the independent components, we need a quantitative measure of non-Gaussianity.
\end{itemize}
\pspread
Consider the random variable $y$ and assume that it has zero mean and unit variance (enforced by pre-processing).\\[.5cm] \pause
We will consider \structure{three measures of non-Gaussianity}:
\begin{itemize}
\item \structure{Kurtosis}
\item \structure{Negentropy}
\item \structure{Mutual Information}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Kurtosis}
\begin{citeblock}{Definition}
The \structure{\textit{Kurtosis}} of $y$ is defined as:
\begin{displaymath}
\kurt(y) = E \{ y^4 \} - 3 \left( E\{ y^2 \} \right)^2
\end{displaymath}
\pause
%
Since $y$ has unit variance, the equation simplifies to:
\begin{displaymath}
\kurt(y) = E \{ y^4 \} - 3
\end{displaymath}
\pause
%
Linearity properties for independent random variables $y_1, y_2$:
\begin{eqnarray*}
\kurt(y_1 + y_2) &=& \kurt(y_1) + \kurt(y_2) \\
\kurt(\alpha y) &=& \alpha^4 \kurt(y) ~, \quad \alpha \in \real
\end{eqnarray*}
\end{citeblock}
\end{frame}
\begin{frame}
\frametitle{Kurtosis \cont}
\structure{Kurtosis for a Gaussian distribution:}
\begin{itemize}
\item The $n$-th central moment of a Gaussian distribution $p(y) = {\mathcal{N}}(y | \mu, \sigma^2)$ with mean $\mu$ and variance $\sigma^2$ is:
\begin{eqnarray*}
E \{ (y - \mu)^n \} = \begin{cases}
(n-1) !! \cdot \sigma^n \quad &, ~n~ \mbox{even} \\
0 \quad &, ~n~ \mbox{odd}
\end{cases}
\end{eqnarray*}
\structure{Note:}
$(n)!!$ denotes the double factorial, i.\,e.\ the product of every odd \\
\hspace{0.85cm} number from 1 to $n$. \pause
\item Thus, for a zero mean, unit variance random variable $y$ that is normally distributed:
\begin{displaymath}
\kurt(y) = 0
\end{displaymath}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Kurtosis \cont}
\structure{Properties of Kurtosis:}
\begin{itemize}
\item Kurtosis is zero for a Gaussian random variable. \pause
\item For most (but not all) non-Gaussian random variables, \\
Kurtosis is nonzero. \pause
\item Kurtosis can be positive or negative. \pause
\item Typically, non-Gaussianity is measured as:
\begin{itemize}
\item \structure{$| \kurt(y) |$} ~ or
\item \structure{$\kurt(y)^2$}
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Kurtosis \cont}
\begin{ovalblock}{Example}
\begin{columns}
\column{.5\textwidth}
\structure{\small \hspace{.5cm} Subgaussian pdf:}
{\footnotesize
\begin{displaymath}
p(y) = \begin{cases}
\frac{1}{2\sqrt{3}} \quad &, ~\mbox{if}~|y| \leq \sqrt{3} \\
0 \quad &, ~\mbox{otherwise}
\end{cases}
\end{displaymath}
}
\column{.5\textwidth}
\structure{\small \hspace{.5cm} Supergaussian pdf:}
{\footnotesize
\begin{displaymath}
p(y) = \frac{1}{\sqrt{2}} \exp(-\sqrt{2} |y|)
\end{displaymath}
}
\end{columns}
\vspace{-0.5cm}
\begin{columns}
\column{.5\textwidth}
\begin{center}
\resizebox{.85\linewidth}{!}{
\input{\texfigdir/subgaussian.pstex_t}
}
\end{center}
%
\vspace{-0.3cm}
{\footnotesize
\begin{displaymath}
\kurt(y) < 0
\end{displaymath}
}
\column{.5\textwidth}
\begin{center}
\resizebox{.85\linewidth}{!}{
\input{\texfigdir/supergaussian.pstex_t}
}
\end{center}
%
\vspace{-0.3cm}
{\footnotesize
\begin{displaymath}
\kurt(y) > 0
\end{displaymath}
}
\end{columns}
\end{ovalblock}
\end{frame}
\begin{frame}
\frametitle{Kurtosis \cont}
\begin{itemize}
\item Consider 2-D case using the linear combination:
\begin{eqnarray*}
y &=& \vec w^T \vec x = \vec w^T \mat A \vec s = \vec z^T \vec s = z_1 s_1 + z_2 s_2 \\
\kurt(y) &=& \kurt(z_1 s_1) + \kurt(z_2 s_2) = z_1^4 \kurt(s_1) + z_2^4 \kurt(s_2) \pause
\end{eqnarray*}
\item As $y$ has unit variance, concerning also $s_1, s_2$:
\begin{displaymath}
E\{ y^2 \} = z_1^2 + z_2^2 = 1
\end{displaymath}
which constrains $\vec z$ to the unit circle in the 2-D plane. \\[.25cm] \pause
\item Thus, we have to find the maximum of the following function on the unit circle w.\,r.\,t.\ $\vec z$:
\begin{displaymath}
| \kurt(y) | = | z_1^4 \kurt(s_1) + z_2^4 \kurt(s_2) |
\end{displaymath}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Kurtosis}
\structure{Optimization landscape for Kurtosis in 2-D plane:}
\begin{center}
\resizebox{.4\linewidth}{!}{
\input{\texfigdir/kurtosis_optimization.pstex_t}
}
\end{center}
\begin{itemize}
\item Thick curve is the unit circle
\item Thin curves are isocontours of the objective function \pause
\item The maxima are located at sparse values of $\vec z$, i.\,e.\ when:
\begin{center}
\tikz[baseline]{
\node[fill=bl1!100, anchor=base, rounded corners=3pt, inner sep=2mm] (d1) {
\color{bl3}
$y = \pm s_i$
};
}
\end{center}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Kurtosis \cont}
\structure{Maximizing the non-Gaussianity of a vector $\vec w$ in practice:}
\begin{itemize}
\item Start with some initial vector $\vec w$. \pause
\item Use a gradient descent method to optimize:
\begin{displaymath}
\argmax_{\vec w} | \kurt (y) | = \argmax_{\vec{w}} | \kurt(\vec w^T \vec x) | \pause
\end{displaymath}
\item Plug this optimization into the ICA estimation algorithm (see above).
\end{itemize}
\end{frame}