-
Notifications
You must be signed in to change notification settings - Fork 33
/
Copy pathlearn-lang-diary-part-eight.lyx
9438 lines (7229 loc) · 175 KB
/
learn-lang-diary-part-eight.lyx
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
#LyX 2.3 created this file. For more info see http://www.lyx.org/
\lyxformat 544
\begin_document
\begin_header
\save_transient_properties true
\origin unavailable
\textclass article
\begin_preamble
\usepackage{url}
\usepackage{slashed}
\end_preamble
\use_default_options false
\maintain_unincluded_children false
\language english
\language_package default
\inputencoding utf8
\fontencoding global
\font_roman "times" "default"
\font_sans "helvet" "default"
\font_typewriter "cmtt" "default"
\font_math "auto" "auto"
\font_default_family default
\use_non_tex_fonts false
\font_sc false
\font_osf false
\font_sf_scale 100 100
\font_tt_scale 100 100
\use_microtype false
\use_dash_ligatures false
\graphics default
\default_output_format default
\output_sync 0
\bibtex_command default
\index_command default
\paperfontsize default
\spacing single
\use_hyperref true
\pdf_bookmarks true
\pdf_bookmarksnumbered false
\pdf_bookmarksopen false
\pdf_bookmarksopenlevel 1
\pdf_breaklinks true
\pdf_pdfborder true
\pdf_colorlinks true
\pdf_backref false
\pdf_pdfusetitle true
\papersize default
\use_geometry false
\use_package amsmath 2
\use_package amssymb 2
\use_package cancel 1
\use_package esint 0
\use_package mathdots 1
\use_package mathtools 1
\use_package mhchem 0
\use_package stackrel 1
\use_package stmaryrd 1
\use_package undertilde 1
\cite_engine basic
\cite_engine_type default
\biblio_style plain
\use_bibtopic false
\use_indices false
\paperorientation portrait
\suppress_date false
\justification true
\use_refstyle 0
\use_minted 0
\index Index
\shortcut idx
\color #008000
\end_index
\secnumdepth 3
\tocdepth 3
\paragraph_separation indent
\paragraph_indentation default
\is_math_indent 0
\math_numbering_side default
\quotes_style english
\dynamic_quotes 0
\papercolumns 1
\papersides 1
\paperpagestyle default
\listings_params "basicstyle={\ttfamily},basewidth={0.45em}"
\tracking_changes false
\output_changes false
\html_math_output 0
\html_css_as_file 0
\html_be_strict false
\end_header
\begin_body
\begin_layout Title
Language Learning Diary - Part Eight
\end_layout
\begin_layout Date
Sept 2022
\end_layout
\begin_layout Author
Linas Vepštas
\end_layout
\begin_layout Abstract
The language-learning effort involves research and software development
to implement the ideas concerning unsupervised learning of grammar, syntax
and semantics from corpora.
This document contains supplementary notes and a loosely-organized semi-chronol
ogical diary of results.
The notes here might not always makes sense; they are a short-hand for
my own benefit, rather than aimed at you, dear reader!
\end_layout
\begin_layout Section*
Introduction
\end_layout
\begin_layout Standard
Part Eight of the diary consists of two parts.
First, a very short collection of notes on hypervectors and Gaussian orthogonal
ensembles.
Collected here as a handy reference, because adequate articles on some
of these ideas do not yet exist.
The second part is an exploration of applying the idea of Gaussian orthogonal
ensembles to a current dataset.
Overall, a tremendous success! This appears to provide an excellent metric
of word-similarity! Victory!
\end_layout
\begin_layout Section*
Summary Conclusions
\end_layout
\begin_layout Standard
A summary of what is found in this part of the diary:
\end_layout
\begin_layout Itemize
The first few sections discuss hypervectors.
These are interesting, but do not appear to be directly useful in the current
situation.
\end_layout
\begin_layout Itemize
The next section is titled
\begin_inset Quotes eld
\end_inset
Gaussian Orthogonal Ensemble
\begin_inset Quotes erd
\end_inset
(but this name is misleading and incorrect.
I don't have a better name yet).
In Chapters Three and Five, it was noted that the symmetric-MI of word
pairs (built up out of disjuncts) is distributed as a Gaussian.
Whenever one has such a situation, the points can be understood to be vectors
on a high-dimensional unit sphere.
In this case, associated to each word
\begin_inset Formula $w$
\end_inset
is a unit-length word-vectors
\begin_inset Formula $\hat{w}$
\end_inset
which is uniformly distributed on the unit sphere
\begin_inset Formula $S_{N-1}$
\end_inset
where
\begin_inset Formula $N$
\end_inset
is the size of the vocabulary.
\end_layout
\begin_layout Itemize
Such a distribution is perfectly uniform, whenever the sampling is taken
from a perfect Gaussian; this is how uniform distributions on spheres are
defined.
\end_layout
\begin_layout Itemize
Given two vectors
\begin_inset Formula $\hat{w}$
\end_inset
and
\begin_inset Formula $\hat{u}$
\end_inset
drawn from a uniform distribution, we expect the angle between them:
\begin_inset Formula $\theta=\arccos\left(\hat{w}\cdot\hat{u}\right)$
\end_inset
to be uniformly distributed on the interval
\begin_inset Formula $\left[0,\pi\right]$
\end_inset
.
\end_layout
\begin_layout Itemize
The actual word-data is approximately uniform.
The current favorite dataset is explored.
We find that the word-vectors
\begin_inset Formula $\hat{w}$
\end_inset
are
\begin_inset Quotes eld
\end_inset
almost
\begin_inset Quotes erd
\end_inset
uniformly distributed, except that:
\end_layout
\begin_deeper
\begin_layout Itemize
There are very few word-pairs for which
\begin_inset Formula $\theta\apprle\pi/6$
\end_inset
and that those which do lie in this region are strongly similar, grammatically.
By visual inspection, the similarity is excellent.
This is the primary, most important result of this diary chapter.
\end_layout
\begin_layout Itemize
There are few or no word-pairs for which
\begin_inset Formula $\theta\gtrsim3\pi/4$
\end_inset
.
This means that there aren't any words which are
\begin_inset Quotes eld
\end_inset
truly grammatically dissimilar from one-another
\begin_inset Quotes erd
\end_inset
.
This result does not seem to be important or usable; just curious.
\end_layout
\end_deeper
\begin_layout Itemize
Given the collection of values
\begin_inset Formula $\theta\left(w,u\right)$
\end_inset
over pairs
\begin_inset Formula $\left(u,w\right)$
\end_inset
, the trick of projecting to a sphere can be repeated, giving another set
of vectors.
This can again be repeated,
\emph on
ad infinitum
\emph default
.
The first repetition is explored.
It looks OK; it provides similarities that seem to be just about as good
as the original sphere set, maybe every so slightly worse.
This second recomputation of similarities requires additional CPU time,
which is considerable, and so does not seem worth the effort.
It's measuring some hard-to-comprehend second-order effect in the distribution
of grammatical similarity.
That is to say, it's physical interpretation is unclear.
\end_layout
\begin_layout Standard
That's it.
Moving forward, it is clear that
\begin_inset Formula $\theta=\arccos\left(\hat{w}\cdot\hat{u}\right)$
\end_inset
is the superior metric for measuring word grammatical similarity.
And BTW, it is a true metric, satisfying the triangle inequality.
It should form the foundation for future clustering work.
\end_layout
\begin_layout Section*
Bipolar Hypervectors
\end_layout
\begin_layout Standard
A bipolar hypervector is a vector in a
\begin_inset Formula $D$
\end_inset
-dimensional space having values in the set
\begin_inset Formula $\mathbb{Z}_{2}=\left\{ -1,1\right\} $
\end_inset
; that is, a vector in
\begin_inset Formula $\left\{ -1,1\right\} ^{D}.$
\end_inset
Every hypervector corresponds to a vertex of a hypercube centered at the
origin of a
\begin_inset Formula $D$
\end_inset
-dimensional space.
Bipolar hypervectors have the interesting property that, for
\begin_inset Formula $D$
\end_inset
even, given any (random) vector
\begin_inset Formula $v$
\end_inset
, if one flips half the bits to get a vector
\begin_inset Formula $w$
\end_inset
, then
\begin_inset Formula $v$
\end_inset
and
\begin_inset Formula $w$
\end_inset
are orthogonal!
\end_layout
\begin_layout Standard
This interesting property can be used to map point sequences (
\begin_inset Quotes eld
\end_inset
curves
\begin_inset Quotes erd
\end_inset
) to sequences of bipolar hypervectors such that the endpoints of the point
sequence are orthogonal, and intermediate points have increasingly larger
cosine distances.
Geometrically, this maps the point sequence to a sequence of corners on
the hypercube that are increasingly distant from the starting point.
of point sequences to hypervectors
\end_layout
\begin_layout Standard
This map is a homomorphism that preserves the metric on the point sequence.
This metric property can then be deployed to simplify classification problems,
by mapping the space to be classified to vector arithmetic and cosine distances.
These two ideas are developed below.
\end_layout
\begin_layout Subsection*
Metric properties
\end_layout
\begin_layout Standard
Let
\begin_inset Formula $\left[p_{0},\cdots,p_{N}\right]$
\end_inset
be a totally ordered sequence of points.
The total order is just
\begin_inset Formula $p_{i}<p_{j}$
\end_inset
for
\begin_inset Formula $i<j$
\end_inset
for integer index
\begin_inset Formula $i,j$
\end_inset
.
This order can be metricized with a metric
\begin_inset Formula $g$
\end_inset
such that
\begin_inset Formula $g\left(p_{i},p_{j}\right)=\left|i-j\right|/N$
\end_inset
.
The metric is normalized written so that the maximum distance is 1.
Pick a dimension
\begin_inset Formula $D>2N$
\end_inset
and conventionally
\begin_inset Formula $D\gg2N$
\end_inset
and an arbitrary initial (random) vector
\begin_inset Formula $v_{0}$
\end_inset
that will correspond to
\begin_inset Formula $p_{0}$
\end_inset
.
Generate a sequence of bipolar hypervectors
\begin_inset Formula $v_{k}$
\end_inset
as follows.
Given
\begin_inset Formula $v_{i}$
\end_inset
, select (randomly)
\begin_inset Formula $D/2N$
\end_inset
bits that have not been selected before, and flip them, to obtain
\begin_inset Formula $v_{i+1}$
\end_inset
.
\end_layout
\begin_layout Standard
The above generates a sequence of (bipolar hyperdimensional) vectors with
the following properties.
The dot product is
\begin_inset Formula
\[
v_{k}\cdot v_{k}=D
\]
\end_inset
For neighboring points, the dot product is
\begin_inset Formula
\[
v_{k}\cdot v_{k+1}=D\left(1-\frac{1}{N}\right)
\]
\end_inset
because these differ in
\begin_inset Formula $D/2N$
\end_inset
bit locations.
(The Hamming distance is
\begin_inset Formula $D/2N$
\end_inset
; so
\begin_inset Formula $D/2N$
\end_inset
bit positions that are
\begin_inset Formula $+$
\end_inset
are replaced by
\begin_inset Formula $-$
\end_inset
, and so that total sum decreases by
\begin_inset Formula $N$
\end_inset
.)
\end_layout
\begin_layout Standard
In general, the Hamming distance between
\begin_inset Formula $v_{k}$
\end_inset
and
\begin_inset Formula $v_{k+n}$
\end_inset
is
\begin_inset Formula $nD/2N$
\end_inset
and so the dot product is
\begin_inset Formula
\[
v_{k}\cdot v_{k+n}=D\left(1-\frac{n}{N}\right)
\]
\end_inset
or equivalently
\begin_inset Formula
\[
v_{i}\cdot v_{j}=D\left(1-\frac{\left|i-j\right|}{N}\right)
\]
\end_inset
so that
\begin_inset Formula
\[
v_{0}\cdot v_{N}=0
\]
\end_inset
are orthogonal.
The Hamming distance between orthogonal vectors is necessarily
\begin_inset Formula $D/2$
\end_inset
.
\end_layout
\begin_layout Standard
Normalizing by
\begin_inset Formula $D$
\end_inset
and subtracting from 1 reproduces the original metric on the point sequence:
i.e.
\begin_inset Formula
\[
g\left(p_{i},p_{j}\right)=\frac{\left|i-j\right|}{N}=1-\frac{v_{i}\cdot v_{j}}{D}
\]
\end_inset
\end_layout
\begin_layout Standard
Of course, all this machination is pointless for one-dimensional point sequences.
So ...
for the more complex case.
\end_layout
\begin_layout Subsection*
Dimensional Oxidation
\end_layout
\begin_layout Standard
In chemistry, oxidation is the opposite of reduction.
If dimensional reduction is the reduction of the number of dimensions to
describe a dataset, then playing on this, dimensional oxidation is the
act of increasing dimensions.
\end_layout
\begin_layout Standard
One conventional machine learning problem is the classification of regions
of some
\begin_inset Formula $M$
\end_inset
-dimensional space features.
That is, there are a set of real-valued features
\begin_inset Formula $f_{m}\in\mathbb{R}$
\end_inset
for
\begin_inset Formula $1\le m\le M$
\end_inset
.
These are presumed to be bounded, so that
\begin_inset Formula $f_{m}^{\mathrm{min}}\le f_{m}\le f_{m}^{\mathrm{max}}$
\end_inset
so that these can be normalized to the unit cube
\begin_inset Formula $\left[0,1\right]^{M}$
\end_inset
by writing
\begin_inset Formula
\[
x_{m}=\frac{f_{m}-f_{m}^{\mathrm{min}}}{f_{m}^{\mathrm{max}}-f_{m}^{\mathrm{min}}}
\]
\end_inset
Each unit interval may be digitized (partitioned) into
\begin_inset Formula $N+1$
\end_inset
distinct sub-intervals.
This partitioning
\begin_inset Formula $\left[p_{0},\cdots,p_{N}\right]$
\end_inset
then provides a totally ordered points sequence that can be mapped to bipolar
hypervectors.
A given point
\begin_inset Formula $x\in\left[0,1\right]^{M}$
\end_inset
is thus mapped to
\begin_inset Formula $M$
\end_inset
vectors
\begin_inset Formula $v_{m}$
\end_inset
.
Summing these provides a mapping of the unit cube to
\begin_inset Formula $\mathbb{Z}^{D}$
\end_inset
:
\begin_inset Formula
\[
w=w\left(x\right)=\sum_{m}v_{m}
\]
\end_inset
Since the sum is bounded between
\begin_inset Formula $-M$
\end_inset
and
\begin_inset Formula $M$
\end_inset
in each direction, this vector lies on the hypercube lattice
\begin_inset Formula $\left(2M\right)^{D}$
\end_inset
.
\end_layout
\begin_layout Standard
If these vectors are normalized to unit length, then they live on the surface
of the hyper-sphere
\begin_inset Formula $S_{D-1}$
\end_inset
.
In general, these points are not random, evenly distributed on the hyper-sphere.
In a narrow sense, we can offer a theorem: the points are randomly distributed
on the hyper-sphere if and only if they are randomly distributed in the
unit cube
\begin_inset Formula $\left[0,1\right]^{M}$
\end_inset
.
However, in a broader sense, when
\begin_inset Formula $N\ll D$
\end_inset
, points are increasingly scattered on the unit sphere, becoming increasingly
uniformly distributed.
This is effectively because the start and end points of the unit interval
are mapped to random hypervectors.
(This follows(?) because random hypervectors have a binomial bit distribution,
and for large
\begin_inset Formula $D$
\end_inset
, the binomial distribution approaches the Gaussian).
\end_layout
\begin_layout Standard
So again, we seem to approach the case of a Gaussian Orthogonal Ensemble.
Kind of...
\end_layout
\begin_layout Subsection*
Efficient Dimensional Oxidation
\end_layout
\begin_layout Standard
A way to encode low-dimensional classification problems into hypervectors
that sharply improves on the naive uniform feature-space digitization is
described in Basaklar,
\emph on
et al.
\emph default
\begin_inset Quotes eld
\end_inset
Hypervector Design for Efficient Hyperdimensional Computing on Edge Devices
\begin_inset Quotes erd
\end_inset
\begin_inset CommandInset href
LatexCommand href
name "https://arxiv.org/pdf/2103.06709.pdf"
target "https://arxiv.org/pdf/2103.06709.pdf"
literal "false"
\end_inset
\end_layout
\begin_layout Standard
The presumption is that there is a preexisting training dataset, the parameter
space is low dimensional, and the number of clusters is fixed.
\end_layout
\begin_layout Standard
The solution is to divide up the parameter space in a non-uniform kind of
way, devoting lots of extra hypervector bit-flips to the boundary zones
between clusters, so that the boundaries of the clusters can be cleanly
distinguished.
This is done by specifying an integer optimization problem.
The trick is to (i) maximize the training accuracy and (ii) minimize the
similarity between class encoders, subject to (iii) orthogonalization of
parameter endpoints.
Because this is an integer optimization problem, a genetic algorithm is
used to perform the search.
The paper provides details.
\end_layout
\begin_layout Standard
This is not directly relevant for us, because we don't have a training set,
nor do we know
\emph on
a priori
\emph default
how many clusters there will be.
\end_layout
\begin_layout Subsection*
Factoids
\end_layout
\begin_layout Standard
Assorted notes:
\end_layout
\begin_layout Itemize
Almost all random vectors are orthogonal to one another (or nearly so).
This follows from binomial coefficients being approximations for Gaussians.
There are
\begin_inset Formula
\[
{D \choose D/2}=\frac{D!}{\left(D/2\right)!^{2}}\approx\sqrt{\frac{2}{\pi D}}\,2^{D}
\]
\end_inset
orthogonal vectors, which follows from Stirling's law
\begin_inset Formula $n!\approx\sqrt{2\pi n}e^{-n}n^{n}$
\end_inset
.
There are
\begin_inset Formula
\[
{D \choose \frac{D}{2}+1}=\frac{D!}{\left(\frac{D}{2}-1\right)!\left(\frac{D}{2}+1\right)!}\approx\sqrt{2/\pi D}\,2^{D}xxx
\]
\end_inset
almost orthogonal vectors, that differ by one bit.
More generally, for
\begin_inset Formula $n\ll D$
\end_inset
there are
\begin_inset Formula
\[
{D \choose \frac{D}{2}+n}=\frac{D!}{\left(\frac{D}{2}-n\right)!\left(\frac{D}{2}+n\right)!}\approx\sqrt{2/\pi D}\,2^{D}\left(1-\frac{2n^{2}}{D}\right)
\]
\end_inset
vectors that differ by
\begin_inset Formula $n$
\end_inset
bits.
(Need to double check this, might be errors).
\end_layout
\begin_layout Itemize
Given two corners on a hypercube differing by
\begin_inset Formula $d$
\end_inset
bits, there are
\begin_inset Formula $2^{d}$
\end_inset
shortest paths between them.
\end_layout
\begin_layout Itemize
The midway points on such paths can have larger Jacquard (Hamming) distances
to each other, than to either endpoint.
In fact, this will almost always be the case.
\end_layout
\begin_layout Itemize
If the vectors are normalized, they can be seen to live on the surface of
a sphere (of the same dimension).
\end_layout
\begin_layout Itemize
Ternary hypervectors i.e.
elements of
\begin_inset Formula $\left\{ -1,0,1\right\} ^{D}$
\end_inset
form a field under point-wise multiplication and sgn applied to arithmetic
addition.
\end_layout
\begin_layout Itemize
The projection of lattice points in
\begin_inset Formula $\mathbb{Z}^{D}$
\end_inset
to the unit sphere
\begin_inset Formula $S_{D-1}$
\end_inset
is presumably dense.
No clue if its
\begin_inset Quotes eld
\end_inset
uniformly
\begin_inset Quotes erd
\end_inset
distributed; presumably its not, much like the rationals in the unit interval.
\end_layout
\begin_layout Itemize
No clue what the analogs of the modular group or the fundamental domain
are.
\end_layout
\begin_layout Standard
\end_layout
\begin_layout Section*
Spin Glasses
\end_layout
\begin_layout Standard
Cribbed notes from Michel TALAGRAND
\begin_inset Quotes eld
\end_inset
Mean Field Models for Spin Glasses Volume I: Basic Examples
\begin_inset Quotes erd
\end_inset
(2010) Springer-Verlag.
\end_layout
\begin_layout Section*
Gaussian Orthogonal Ensembles
\end_layout
\begin_layout Standard
Well, not really; that's just the working title.
There's no actual ensemble.
\end_layout
\begin_layout Standard
The thing to explore is this: suppose there's an index
\begin_inset Formula $w,u\in\left\{ 1,\cdots,N\right\} $
\end_inset
i.e.
\begin_inset Formula $N$
\end_inset
-dimensional.
Suppose there are numbers
\begin_inset Formula $f\left(w,u\right)=f\left(u,w\right)$
\end_inset
which are distributed with a normal distribution with mean
\begin_inset Formula $\mu$
\end_inset
and stddev
\begin_inset Formula $\sigma$
\end_inset
.
Experimentally, we compute
\begin_inset Formula
\[
\mu=\frac{1}{N^{2}}\sum_{u,w}f\left(u,w\right)=\left\langle f\right\rangle
\]
\end_inset
and
\begin_inset Formula
\begin{align*}
\sigma^{2}= & \frac{1}{N^{2}}\sum_{u,w}\left(f\left(u,w\right)-\mu\right)^{2}\\
= & \frac{1}{N^{2}}\sum_{u,w}f^{2}\left(u,w\right)-\mu^{2}\\
= & \left\langle f^{2}\right\rangle -\left\langle f\right\rangle ^{2}
\end{align*}
\end_inset
as usual.
Define
\begin_inset Formula $g\left(u,w\right)=\left(f\left(u,w\right)-\mu\right)/\sigma$
\end_inset
and then
\begin_inset Formula $g\left(u,w\right)$
\end_inset
is normally distributed about zero with unit stddev.
\end_layout
\begin_layout Standard
Fix
\begin_inset Formula $w$
\end_inset
and define an
\begin_inset Formula $N$
\end_inset
-dimensional vector
\begin_inset Formula $\vec{w}\in\mathbb{R}^{N}$
\end_inset
whose vector components are
\begin_inset Formula $w_{u}=g\left(w,u\right)$
\end_inset
.
Normalize them to unit length, so that
\begin_inset Formula $\hat{w}=\vec{w}/\left\Vert \vec{w}\right\Vert $
\end_inset
where
\begin_inset Formula $\left\Vert \vec{w}\right\Vert =\left\Vert \vec{w}\right\Vert _{2}$
\end_inset
is the Euclidean norm.
This is the Gaussian orthogonal ensemble.
There are
\begin_inset Formula $N$
\end_inset
of these vectors, they are uniformly distributed on the
\begin_inset Formula $N-1$
\end_inset
sphere
\begin_inset Formula $S_{N-1}$
\end_inset
.
\end_layout
\begin_layout Standard
The experimental goal is to obtain these vectors, on the actual datasets,
were
\begin_inset Formula $f\left(w,u\right)$
\end_inset
is the MI for two words, the MI being given via the disjunct+shape formulas
explored in earlier chapters.
So lets see what we get.
\end_layout
\begin_layout Subsection*
Ranking
\end_layout
\begin_layout Standard
The ranked MI has the form
\begin_inset Formula $f\left(w,u\right)=s\left(w,u\right)-r\left(w\right)-r\left(u\right)$
\end_inset
.