目錄
第1 章簡介..................................................................1
1.1 視覺問答的動機........................................................1
1.2 人工智能任務中的視覺問答...........................................4
1.3 視覺問答類別..........................................................5
1.3.1 數(shù)據(jù)分類驅(qū)動......................................................6
1.3.2 任務分類驅(qū)動......................................................7
1.3.3 其他..............................................................7
參考文獻.....................................................................8

第1 部分基礎(chǔ)理論
第2 章深度學習基礎(chǔ)......................................................15
2.1 神經(jīng)網(wǎng)絡...............................................................15
2.2 卷積神經(jīng)網(wǎng)絡..........................................................17
2.3 循環(huán)神經(jīng)網(wǎng)絡及變體...................................................18
2.4 編碼器-解碼器結(jié)構(gòu)....................................................20
2.5 注意力機制.............................................................20
2.6 記憶網(wǎng)絡...............................................................21
2.7 Transformer 網(wǎng)絡和BERT............................................22
2.8 圖神經(jīng)網(wǎng)絡.............................................................24
參考文獻.....................................................................25

第3 章問答基礎(chǔ)知識......................................................27
3.1 基于規(guī)則的方法........................................................27
3.2 基于信息檢索的方法...................................................28
3.3 問答的神經(jīng)語義解析...................................................29
3.4 問答知識庫.............................................................29
參考文獻.....................................................................30

第2 部分圖像視覺問答
第4 章經(jīng)典視覺問答......................................................35
4.1 簡介....................................................................35
4.2 數(shù)據(jù)集..................................................................36
4.3 生成與分類：兩種回答策略...........................................40
4.4 聯(lián)合嵌入...............................................................40
4.4.1 序列到序列編碼器-解碼器模型......................................40
4.4.2 雙線性編碼模型....................................................43
4.5 注意力機制.............................................................45
4.5.1 堆疊注意力網(wǎng)絡....................................................45
4.5.2 分層問題-圖像協(xié)同注意力..........................................47
4.5.3 自底向上和自頂向下的注意力.......................................49
4.6 記憶網(wǎng)絡...............................................................51
4.6.1 改進的動態(tài)記憶網(wǎng)絡...............................................51
4.6.2 記憶增強網(wǎng)絡......................................................52
4.7 組合推理...............................................................54
4.7.1 神經(jīng)模塊網(wǎng)絡......................................................55
4.7.2 動態(tài)神經(jīng)模塊網(wǎng)絡..................................................56
4.8 圖神經(jīng)網(wǎng)絡.............................................................58
4.8.1 圖卷積網(wǎng)絡........................................................58
4.8.2 圖注意力網(wǎng)絡......................................................60
4.8.3 視覺問答圖卷積網(wǎng)絡...............................................62
4.8.4 視覺問答圖注意力網(wǎng)絡.............................................63
參考文獻.....................................................................66

第5 章基于知識的視覺問答..............................................71
5.1 簡介....................................................................71
5.2 數(shù)據(jù)集..................................................................72
5.3 知識庫..................................................................74
5.3.1 數(shù)據(jù)庫百科........................................................74
5.3.2 ConceptNet........................................................74
5.4 知識嵌入...............................................................75
5.4.1 文字對矢量表示法..................................................75
5.4.2 基于BERT 的表征.................................................78
5.5 問題-查詢轉(zhuǎn)換.........................................................79
5.5.1 基于查詢映射的方法...............................................79
5.5.2 基于學習的方法....................................................81
5.6 查詢知識庫的方法.....................................................82
5.6.1 RDF ..............................................................82
5.6.2 記憶網(wǎng)查詢........................................................83
參考文獻.....................................................................84

第6 章視覺問答的視覺和語言預訓練..................................88
6.1 簡介....................................................................88
6.2 常規(guī)預訓練模型........................................................89
6.2.1 ELMo .............................................................89
6.2.2 GPT ..............................................................89
6.2.3 MLM .............................................................90
6.3 視覺和語言預訓練的常用方法.........................................93
6.3.1 單流方法..........................................................94
6.3.2 雙流方法..........................................................96
6.4 視覺問答及其下游任務微調(diào)...........................................98
參考文獻.....................................................................101

第3 部分視頻視覺問答
第7 章視頻表征學習.....................................................·105
7.1 人工標注的局部視頻描述符...........................................105
7.2 數(shù)據(jù)驅(qū)動的深度學習的視頻特征表示.................................107
7.3 視頻表征的自監(jiān)督學習................................................109
參考文獻.....................................................................110

第8 章視頻問答...........................................................·112
8.1 簡介....................................................................112
8.2 數(shù)據(jù)集..................................................................112
8.2.1 多步推理數(shù)據(jù)集....................................................113
8.2.2 單步推理數(shù)據(jù)集....................................................116
8.3 使用編碼器-解碼器結(jié)構(gòu)的傳統(tǒng)視頻時空推理.........................118
參考文獻.....................................................................123

第9 章視頻問答的高級模型.............................................·126
9.1 時空特征注意力........................................................126
9.2 記憶網(wǎng)絡...............................................................129
9.3 時空圖神經(jīng)網(wǎng)絡........................................................130
參考文獻.....................................................................132

第4 部分視覺問答高級任務
第10 章具身視覺問答...................................................·137
10.1 簡介...................................................................137
10.2 模擬器、數(shù)據(jù)集和評估標準..........................................138
10.2.1 模擬器...........................................................138
10.2.2 數(shù)據(jù)集...........................................................140
10.2.3 評估指標.........................................................141
10.3 語言引導的視覺導航.................................................142
10.3.1 視覺和語言導航...................................................142
10.3.2 遠程對象定位.....................................................147
10.4 具身問答..............................................................148
10.5 交互式問答............................................................150
參考文獻.....................................................................151

第11 章醫(yī)學視覺問答...................................................·153
11.1 簡介...................................................................153
11.2 數(shù)據(jù)集.................................................................154
11.3 醫(yī)學視覺問答的經(jīng)典方法............................................156
11.4 醫(yī)學視覺問答的元學習方法..........................................159
11.5 基于BERT 的醫(yī)學視覺問答方法....................................160
參考文獻.....................................................................162

第12 章基于文本的視覺問答...........................................·165
12.1 簡介...................................................................165
12.2 數(shù)據(jù)集.................................................................166
12.2.1 TextVQA.........................................................166
12.2.2 ST-VQA .........................................................167
12.2.3 OCR-VQA .......................................................168
12.3 OCR 標記表示........................................................168
12.4 簡單融合模型.........................................................169
12.5 基于Transformer 的模型............................................170
12.6 圖模型.................................................................172
參考文獻.....................................................................173

第13 章視覺問題生成...................................................·176
13.1 簡介...................................................................176
13.2 數(shù)據(jù)融合中的視覺問題生成..........................................176
13.2.1 從答案生成問題...................................................177
13.2.2 從圖像生成問題...................................................178
13.2.3 對抗學習.........................................................179
13.3 作為視覺理解問題的視覺問題生成..................................180
參考文獻.....................................................................182

第14 章視覺對話.........................................................·185
14.1 簡介...................................................................185
14.2 數(shù)據(jù)集.................................................................186
14.3 注意力機制............................................................187
14.3.1 具有注意力的分層循環(huán)編碼器和記憶網(wǎng)絡...........................187
14.3.2 歷史條件圖像注意力編碼器........................................188
14.3.3 序列協(xié)同注意力生成模型..........................................190
14.3.4 協(xié)同網(wǎng)絡.........................................................192
14.4 視覺指代表達理解....................................................194
14.5 基于圖的方法.........................................................195
14.5.1 視覺表示的場景圖................................................196
14.5.2 用于視覺和對話表示的圖卷積網(wǎng)絡.................................197
14.6 預訓練模型............................................................199
14.6.1 VD-BERT ........................................................200
14.6.2 Visual-Dialog BERT ..............................................201
參考文獻.....................................................................202

第15 章指代表達理解...................................................·204
15.1 簡介...................................................................204
15.2 數(shù)據(jù)集.................................................................205
15.3 二階段模型............................................................206
15.3.1 聯(lián)合嵌入.........................................................206
15.3.2 協(xié)同注意力模型...................................................208
15.3.3 圖模型...........................................................209
15.4 一階段模型............................................................211
15.5 推理過程理解.........................................................212
參考文獻.....................................................................213

第5 部分總結(jié)與展望
第16 章總結(jié)與展望......................................................·219
16.1 總結(jié)...................................................................219
16.2 展望...................................................................219
16.2.1 視覺問答的可解釋性..............................................219
16.2.2 消除偏見.........................................................220
16.2.3 附加設(shè)置及應用...................................................221
參考文獻.....................................................................221