-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
1460 lines (1000 loc) · 62 KB
/
index.html
File metadata and controls
1460 lines (1000 loc) · 62 KB
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
<!DOCTYPE html>
<html lang="zh-Hans">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=2">
<meta name="theme-color" content="#222">
<meta name="generator" content="Hexo 6.1.0">
<link rel="apple-touch-icon" sizes="180x180" href="/images/apple-touch-icon-next.png">
<link rel="icon" type="image/png" sizes="32x32" href="/images/favicon-32x32-next.png">
<link rel="icon" type="image/png" sizes="16x16" href="/images/favicon-16x16-next.png">
<link rel="mask-icon" href="/images/logo.svg" color="#222">
<link rel="stylesheet" href="/css/main.css">
<link rel="stylesheet" href="/lib/font-awesome/css/all.min.css">
<script id="hexo-configurations">
var NexT = window.NexT || {};
var CONFIG = {"hostname":"zerolxf.github.io","root":"/","scheme":"Muse","version":"7.8.0","exturl":false,"sidebar":{"position":"left","display":"post","padding":18,"offset":12,"onmobile":false},"copycode":{"enable":false,"show_result":false,"style":null},"back2top":{"enable":true,"sidebar":false,"scrollpercent":false},"bookmark":{"enable":false,"color":"#222","save":"auto"},"fancybox":false,"mediumzoom":false,"lazyload":false,"pangu":false,"comments":{"style":"tabs","active":null,"storage":true,"lazyload":false,"nav":null},"algolia":{"hits":{"per_page":10},"labels":{"input_placeholder":"Search for Posts","hits_empty":"We didn't find any results for the search: ${query}","hits_stats":"${hits} results found in ${time} ms"}},"localsearch":{"enable":true,"trigger":"auto","top_n_per_article":1,"unescape":false,"preload":false},"motion":{"enable":true,"async":false,"transition":{"post_block":"fadeIn","post_header":"slideDownIn","post_body":"slideDownIn","coll_header":"slideLeftIn","sidebar":"slideUpIn"}},"path":"search.xml"};
</script>
<meta name="description" content="咸鱼也有咸鱼的生活">
<meta property="og:type" content="website">
<meta property="og:title" content="BigHead">
<meta property="og:url" content="http://zerolxf.github.io/index.html">
<meta property="og:site_name" content="BigHead">
<meta property="og:description" content="咸鱼也有咸鱼的生活">
<meta property="og:locale">
<meta property="article:author" content="BigHead">
<meta name="twitter:card" content="summary">
<link rel="canonical" href="http://zerolxf.github.io/">
<script id="page-configurations">
// https://hexo.io/docs/variables.html
CONFIG.page = {
sidebar: "",
isHome : true,
isPost : false,
lang : 'zh-Hans'
};
</script>
<title>BigHead</title>
<noscript>
<style>
.use-motion .brand,
.use-motion .menu-item,
.sidebar-inner,
.use-motion .post-block,
.use-motion .pagination,
.use-motion .comments,
.use-motion .post-header,
.use-motion .post-body,
.use-motion .collection-header { opacity: initial; }
.use-motion .site-title,
.use-motion .site-subtitle {
opacity: initial;
top: initial;
}
.use-motion .logo-line-before i { left: initial; }
.use-motion .logo-line-after i { right: initial; }
</style>
</noscript>
</head>
<body itemscope itemtype="http://schema.org/WebPage">
<div class="container use-motion">
<div class="headband"></div>
<header class="header" itemscope itemtype="http://schema.org/WPHeader">
<div class="header-inner"><div class="site-brand-container">
<div class="site-nav-toggle">
<div class="toggle" aria-label="Toggle navigation bar">
<span class="toggle-line toggle-line-first"></span>
<span class="toggle-line toggle-line-middle"></span>
<span class="toggle-line toggle-line-last"></span>
</div>
</div>
<div class="site-meta">
<a href="/" class="brand" rel="start">
<span class="logo-line-before"><i></i></span>
<h1 class="site-title">BigHead</h1>
<span class="logo-line-after"><i></i></span>
</a>
</div>
<div class="site-nav-right">
<div class="toggle popup-trigger">
<i class="fa fa-search fa-fw fa-lg"></i>
</div>
</div>
</div>
<nav class="site-nav">
<ul id="menu" class="main-menu menu">
<li class="menu-item menu-item-home">
<a href="/" rel="section"><i class="fa fa-home fa-fw"></i>Home</a>
</li>
<li class="menu-item menu-item-about">
<a href="/about/" rel="section"><i class="fa fa-user fa-fw"></i>About</a>
</li>
<li class="menu-item menu-item-tags">
<a href="/tags/" rel="section"><i class="fa fa-tags fa-fw"></i>Tags</a>
</li>
<li class="menu-item menu-item-categories">
<a href="/categories/" rel="section"><i class="fa fa-th fa-fw"></i>Categories</a>
</li>
<li class="menu-item menu-item-archives">
<a href="/archives/" rel="section"><i class="fa fa-archive fa-fw"></i>Archives</a>
</li>
<li class="menu-item menu-item-search">
<a role="button" class="popup-trigger"><i class="fa fa-search fa-fw"></i>Search
</a>
</li>
</ul>
</nav>
<div class="search-pop-overlay">
<div class="popup search-popup">
<div class="search-header">
<span class="search-icon">
<i class="fa fa-search"></i>
</span>
<div class="search-input-container">
<input autocomplete="off" autocapitalize="off"
placeholder="Searching..." spellcheck="false"
type="search" class="search-input">
</div>
<span class="popup-btn-close">
<i class="fa fa-times-circle"></i>
</span>
</div>
<div id="search-result">
<div id="no-result">
<i class="fa fa-spinner fa-pulse fa-5x fa-fw"></i>
</div>
</div>
</div>
</div>
</div>
</header>
<div class="back-to-top">
<i class="fa fa-arrow-up"></i>
<span>0%</span>
</div>
<main class="main">
<div class="main-inner">
<div class="content-wrap">
<div class="content index posts-expand">
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/BERT-of-Theseus/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/BERT-of-Theseus/" class="post-title-link" itemprop="url">BERT-of-Theseus</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:01:33" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/BERT-of-Theseus/" class="post-meta-item leancloud_visitors" data-flag-title="BERT-of-Theseus" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/BERT-of-Theseus/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/BERT-of-Theseus/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="BERT-of-Theseus"><a href="#BERT-of-Theseus" class="headerlink" title="BERT-of-Theseus"></a>BERT-of-Theseus</h1><p>论文标题: BERT-of-Theseus: Compressing BERT by Progressive Module Replacing<br>作者: Canwen Xu, Wangchunshu Zhou, Tao Ge, Furu Wei, Ming Zhou<br>时间: 2020<br>核心算法名: BERT-of-Theseus<br>论文链接: <a target="_blank" rel="noopener" href="https://arxiv.org/abs/2002.02925">https://arxiv.org/abs/2002.02925</a><br>参考: <a target="_blank" rel="noopener" href="https://chat.openai.com/c/4b9f76af-42bc-47a4-90e6-cfe71e980e0c">科学空间</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>随着深度学习的普及，许多大型神经模型被提出并在各个领域取得了最先进的性能。特别是在自然语言处理（NLP）中，预训练和微调已经成为大多数任务的新规范。基于Transformer的预训练模型在自然语言理解（NLU）和自然语言生成（NLG）领域都占据了主导地位。这些模型受益于其“超参数化”的特性，并包含数百万甚至数十亿的参数，使其在计算成本和效率方面都显得过于昂贵和低效。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>为了解决这个问题，论文提出了一种新的模型压缩方法，通过逐步替换模块来有效地压缩BERT。该方法首先将原始的BERT划分为几个模块，并构建它们的紧凑替代品。然后，我们随机地用它们的替代品替换原始模块来训练紧凑模块以模仿原始模块的行为。我们逐渐增加替换的概率。这种方式使我们的方法在原始和紧凑模型之间带来了更深层次的交互。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/BERT-of-Theseus/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/" class="post-title-link" itemprop="url">ELECTRA (Stanford University & Google Brain)</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:01:33" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/" class="post-meta-item leancloud_visitors" data-flag-title="ELECTRA (Stanford University & Google Brain)" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="ELECTRA-Stanford-University-amp-Google-Brain"><a href="#ELECTRA-Stanford-University-amp-Google-Brain" class="headerlink" title="ELECTRA (Stanford University & Google Brain)"></a>ELECTRA (Stanford University & Google Brain)</h1><p>论文标题: ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators, 作者: Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning, 时间: 2020, 核心算法名: ELECTRA, 论文链接: <a target="_blank" rel="noopener" href="https://arxiv.org/abs/2003.10555">ELECTRA</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>当前的语言表示学习方法可以被视为学习去噪自动编码器。他们选择未标记输入序列的一小部分（通常为15%），遮蔽这些token的身份（例如，BERT）或对这些token的注意力（例如，XLNet），然后训练网络恢复原始输入。虽然这些遮蔽语言建模（MLM）方法由于学习双向表示而比传统的语言模型预训练更有效，但它们通常需要大量的计算成本，因为网络每个示例只从15%的token中学习。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>作为替代，作者提出了一个更有效的预训练任务，称为替换token检测。与其遮蔽输入，作者的方法通过用从小型生成器网络采样的合理替代品替换一些token来破坏它。然后，作者训练了一个判别模型，该模型预测在被破坏的输入中的每个token是否被生成器样本替换。这种新的预训练任务比MLM更有效，因为该任务是在所有输入token上定义的，而不仅仅是被遮蔽的一小部分。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/ELECTRA%20(Stanford%20University%20&%20Google%20Brain)/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/Linear%20Attention/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/Linear%20Attention/" class="post-title-link" itemprop="url">Linear Attention</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:01:33" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/Linear%20Attention/" class="post-meta-item leancloud_visitors" data-flag-title="Linear Attention" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/Linear%20Attention/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/Linear%20Attention/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="Linear-Attention"><a href="#Linear-Attention" class="headerlink" title="Linear Attention"></a>Linear Attention</h1><p>链接：<a target="_blank" rel="noopener" href="https://spaces.ac.cn/archives/7546">https://spaces.ac.cn/archives/7546</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>Attention机制在自然语言处理（NLP）和计算机视觉（CV）中得到了广泛的应用。然而，标准的Attention机制的计算复杂度为O(n^2)，这在处理长序列时会带来问题。因此，研究者们提出了各种方法来降低Attention的计算复杂度，包括稀疏Attention、Reformer、Linformer等。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>本文主要探讨了如何通过去掉Softmax来实现线性Attention，即将Attention的计算复杂度降低到O(n)。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/Linear%20Attention/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/" class="post-title-link" itemprop="url">Map Net-北大</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:01:26" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/" class="post-meta-item leancloud_visitors" data-flag-title="Map Net-北大" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="Map-Net-北大"><a href="#Map-Net-北大" class="headerlink" title="Map Net-北大"></a>Map Net-北大</h1><h1 id="Variance-reduced-Language-Pretraining-via-a-Mask-Proposal-Network"><a href="#Variance-reduced-Language-Pretraining-via-a-Mask-Proposal-Network" class="headerlink" title="Variance-reduced Language Pretraining via a Mask Proposal Network"></a>Variance-reduced Language Pretraining via a Mask Proposal Network</h1><p>这篇论文的标题是”Variance-reduced Language Pretraining via a Mask Proposal Network”，作者是Liang Chen，发表于2020年。</p>
<h2 id="背景："><a href="#背景：" class="headerlink" title="背景："></a>背景：</h2><p>自我监督学习，也称为预训练，对自然语言处理非常重要。大多数预训练方法首先随机掩盖句子中的一些位置，然后训练模型恢复被掩盖位置的标记。这样，模型可以在没有人工标注的情况下进行训练，并且可以使用大量的数据和数十亿的参数。因此，优化效率变得至关重要。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>本文从梯度方差降低的角度解决这个问题。特别是，首先提出了一个原则性的梯度方差分解定理，该定理显示，语言预训练的随机梯度的方差可以自然地分解为两个项：一项是来自批处理中数据样本的方差，另一项是来自掩码采样的方差。为了减少第二部分的方差，我们利用重要性采样策略，该策略旨在根据提议分布而不是均匀分布来采样掩码。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/Map%20Net-%E5%8C%97%E5%A4%A7/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/" class="post-title-link" itemprop="url">RealFormer-谷歌</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 04:59:56" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/" class="post-meta-item leancloud_visitors" data-flag-title="RealFormer-谷歌" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="RealFormer-谷歌"><a href="#RealFormer-谷歌" class="headerlink" title="RealFormer-谷歌"></a>RealFormer-谷歌</h1><p>论文标题: RealFormer: Transformer Likes Residual Attention, 作者: Ruining He, Anirudh Ravula, Bhargav Kanagal, Joshua Ainslie, 时间: 2021年9月10日, 核心算法名: RealFormer, 论文链接: <a target="_blank" rel="noopener" href="https://arxiv.org/abs/2012.11747">arXiv:2012.11747</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>Transformer是现代自然语言处理（NLP）模型的基础。然而，尽管Transformer在许多任务中表现出色，但其设计和实现仍有改进的空间。本文提出了一种名为RealFormer的新技术，该技术通过在Transformer网络中创建残差注意层（Residual Attention Layer）来改进Transformer的性能。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>RealFormer的目标是改进Transformer网络的性能，使其在各种任务中表现更好。这些任务包括Masked Language Modeling、GLUE、SQuAD、Neural Machine Translation、WikiHop、HotpotQA、Natural Questions和OpenKP等。此外，RealFormer还旨在稳定训练过程，并使模型的注意力更加稀疏。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/RealFormer-%E8%B0%B7%E6%AD%8C/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/" class="post-title-link" itemprop="url">Performer - Google</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:00:40" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/" class="post-meta-item leancloud_visitors" data-flag-title="Performer - Google" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="Performer-Google-amp-University-of-Cambridge"><a href="#Performer-Google-amp-University-of-Cambridge" class="headerlink" title="Performer - Google & University of Cambridge"></a>Performer - Google & University of Cambridge</h1><p>论文标题：Rethinking Attention with Performers<br>作者：Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, David Belanger, Lucy Colwell, Adrian Weller<br>时间：2021<br>核心算法名：Performer<br>论文链接：<a target="_blank" rel="noopener" href="https://arxiv.org/abs/2009.14794">arXiv:2009.14794v4</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>Transformer架构在机器学习的多个领域中都取得了最先进的结果，包括自然语言处理、神经机器翻译、文档生成/摘要、时间序列预测、生成建模（如图像生成）、音乐生成和生物信息学等。然而，Transformer的计算复杂度随着输入序列的长度呈二次增长，这对于处理大规模序列的任务来说是不可接受的。为了解决这个问题，研究者们提出了多种方法，如限制注意力机制只关注局部邻域，或者引入稀疏性、池化压缩、聚类/分箱/卷积技术等结构先验。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>Performer是一种新的Transformer架构，它能够以可证明的准确性估计常规（softmax）全秩注意力Transformer，但只使用线性（而不是二次）的空间和时间复杂度，而且不依赖于任何先验知识，如稀疏性或低秩性。为了近似softmax注意力核，Performer使用了一种新的通过正交随机特征进行快速注意力（FAVOR+）的方法，这可能对可扩展的核方法具有独立的兴趣。FAVOR+还可以用于有效地模拟超出softmax的可核化注意力机制。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/Performer%20-%20Google%20&%20University%20of%20Cambridge/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/" class="post-title-link" itemprop="url">Synthesizer Rethinking Self-Attention for Transformer Models (Google Research)</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:03:03" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/" class="post-meta-item leancloud_visitors" data-flag-title="Synthesizer Rethinking Self-Attention for Transformer Models (Google Research)" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="Synthesizer-Rethinking-Self-Attention-for-Transformer-Models-Google-Research"><a href="#Synthesizer-Rethinking-Self-Attention-for-Transformer-Models-Google-Research" class="headerlink" title="Synthesizer: Rethinking Self-Attention for Transformer Models (Google Research)"></a>Synthesizer: Rethinking Self-Attention for Transformer Models (Google Research)</h1><p>论文标题: Synthesizer: Rethinking Self-Attention for Transformer Models, 作者: Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, Che Zheng, 时间: 2020, 核心算法名: Synthesizer, 论文链接: <a target="_blank" rel="noopener" href="https://arxiv.org/pdf/2005.00743.pdf">https://arxiv.org/pdf/2005.00743.pdf</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>Transformer模型（Vaswani等人，2017）在各种任务中都表现出了成功，这使得Transformer在近年来大大取代了一度流行的自回归和循环模型。Transformer模型的核心是查询-键-值点积注意力。Transformer模型的成功广泛地归因于这种自注意力机制，因为全连接的token图能够模型长距离依赖性，提供了强大的归纳偏见。但是，这篇论文质疑点积自注意力的真正重要性，并通过大量实验发现，随机对齐矩阵的表现出奇地好，而从token-token（查询-键）交互中学习注意力权重虽然有用，但并不那么重要。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>为了解决这个问题，作者提出了Synthesizer，一种无需token-token交互就能学习合成注意力权重的模型。实验结果显示，简单的Synthesizer在与vanilla Transformer模型的比较中，在一系列任务上都取得了高度竞争性的性能，包括机器翻译、语言建模、文本生成和GLUE/SuperGLUE基准测试。当与点积注意力组合时，作者发现Synthesizer始终优于Transformers。此外，作者还对Synthesizer与Dynamic Convolutions进行了额外的比较，结果显示，简单的Random Synthesizer不仅比Dynamic Convolutions快60%，而且还将困惑度提高了相对3.5%。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/Synthesizer:%20Rethinking%20Self-Attention%20for%20Transformer%20Models%20(Google%20Research)/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/" class="post-title-link" itemprop="url">T-TA (Transformer-based Text Autoencoder)</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Created: 2023-08-08 00:00:00 / Modified: 05:01:33" itemprop="dateCreated datePublished" datetime="2023-08-08T00:00:00+08:00">2023-08-08</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">In</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E5%A4%A7%E6%A8%A1%E5%9E%8B/" itemprop="url" rel="index"><span itemprop="name">大模型</span></a>
</span>
</span>
<span id="/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/" class="post-meta-item leancloud_visitors" data-flag-title="T-TA (Transformer-based Text Autoencoder)" title="Views">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">Views: </span>
<span class="leancloud-visitors-count"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine: </span>
<a title="valine" href="/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/" itemprop="commentCount"></span>
</a>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h1 id="T-TA-Transformer-based-Text-Autoencoder"><a href="#T-TA-Transformer-based-Text-Autoencoder" class="headerlink" title="T-TA (Transformer-based Text Autoencoder)"></a>T-TA (Transformer-based Text Autoencoder)</h1><p>论文标题: Fast and Accurate Deep Bidirectional Language Representations for Unsupervised Learning, 作者: Joongbo Shin, Yoonhyung Lee, Seunghyun Yoon, Kyomin Jung, 时间: 2020年4月, 核心算法名: Transformer-based Text Autoencoder (T-TA), 论文链接: <a target="_blank" rel="noopener" href="https://arxiv.org/abs/2004.08097">https://arxiv.org/abs/2004.08097</a></p>
<h2 id="背景"><a href="#背景" class="headerlink" title="背景"></a>背景</h2><p>尽管BERT在各种监督学习任务中取得了成功的性能提升，但将BERT应用于无监督任务仍然存在一个限制，即它需要重复推理以计算上下文语言表示。为了解决这个限制，我们提出了一种名为Transformer-based Text Autoencoder (T-TA)的新型深度双向语言模型。T-TA可以无重复地计算上下文语言表示，并具有像BERT一样的深度双向架构的优点。</p>
<h2 id="解决问题"><a href="#解决问题" class="headerlink" title="解决问题"></a>解决问题</h2><p>在CPU环境的运行时实验中，我们提出的T-TA在重排任务中的性能比基于BERT的模型快6倍多，在语义相似性任务中快12倍。此外，T-TA在上述任务中显示出与BERT相比的竞争甚至更好的准确性。</p>
<!--noindex-->
<div class="post-button">
<a class="btn" href="/2023/08/08/T-TA%20(Transformer-based%20Text%20Autoencoder)/#more" rel="contents">
Read more »
</a>
</div>
<!--/noindex-->
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</article>
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-Hans">
<link itemprop="mainEntityOfPage" href="http://zerolxf.github.io/2023/08/08/UNILM%20/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="BigHead">
<meta itemprop="description" content="咸鱼也有咸鱼的生活">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="BigHead">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">
<a href="/2023/08/08/UNILM%20/" class="post-title-link" itemprop="url">UNILM Unified Pre-training for Language Understanding and Generation</a>
</h2>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">Posted on</span>