深度學(xué)習(xí)

目錄
第 1章淺層模型 ............................................................... 1
1.1深度學(xué)習(xí)史前發(fā)展史 ........................................................ 1
1.1.1數(shù)據(jù)分析的初級階段：手工演算階段 ........................................ 1
1.1.2數(shù)據(jù)分析的中級階段：淺層模型階段 ......................................... 3
1.1.3數(shù)據(jù)分析的高級階段：大數(shù)據(jù)深度學(xué)習(xí)階段 .................................. 4
1.2線性回歸模型 .............................................................. 7
1.2.1極小二乘線性回歸 ........................................................ 8
1.2.2模型選擇：模型容量與過擬合和欠擬合問題 ................................ 19
1.2.3屬性空間、假設(shè)函數(shù)空間與基于核函數(shù)的特征映射 ...................... 22
1.2.4特征選擇 ............................................................... 26
1.2.5回歸分析的概率解釋 ..................................................... 32
1.3 Logistics二分類模型 ..................................................... 33
1.4 Softmax多分類模型 ....................................................... 37
1.5 廣義線性模型 ............................................................ 42
參考文獻(xiàn) .................................................................... 52

第 2章深度神經(jīng)網(wǎng)絡(luò) .......................................................... 56
2.1引言 ..................................................................... 56
2.2 BP神經(jīng)網(wǎng)絡(luò) .............................................................. 58
2.2.1從生物神經(jīng)元到 MP神經(jīng)元模型 .......................................... 58
2.2.2 BP神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu) ........................................................ 59
2.2.3 BP算法 ................................................................ 64
2.2.4 BP算法算例 ............................................................ 72
2.3從 BP網(wǎng)絡(luò)到深度網(wǎng)絡(luò) ...................................................... 73
2.3.1梯度消失：多層 BP網(wǎng)絡(luò)訓(xùn)練難題 ........................................... 74
2.3.2逐層貪心預(yù)訓(xùn)練 +全局微調(diào) ............................................. 76
2.4深度網(wǎng)絡(luò)的進(jìn)一步解釋 ..................................................... 83
2.5克服過擬合：深度網(wǎng)絡(luò)中的正則化技術(shù) ........................................ 85
2.5.1模型約束技術(shù) ........................................................... 86
2.5.2輸入約束技術(shù) ........................................................... 88
2.5.3模型集成技術(shù) ........................................................... 92
2.6深度網(wǎng)絡(luò)發(fā)展史 ........................................................... 98
2.6.1早期神經(jīng)網(wǎng)絡(luò)模型（ 1958—1969年） ....................................... 98
2.6.2深度學(xué)習(xí)萌芽期（ 1969—2006年） .......................................... 99
2.6.3深度學(xué)習(xí)爆發(fā)期（2006年至今） ............................................100
參考文獻(xiàn) ....................................................................101

第 3章卷積神經(jīng)網(wǎng)絡(luò) .........................................................108
3.1引言 ....................................................................108
3.2卷積的數(shù)學(xué)公式及其含義 .................................................109
3.3卷積神經(jīng)網(wǎng)絡(luò)的技術(shù)細(xì)節(jié) .................................................114
3.3.1計算機“眼”中的圖像 ...............................................115
3.3.2卷積神經(jīng)網(wǎng)絡(luò) ..........................................................116
3.3.3卷積神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu) .................................................116
3.3.4 CNNs訓(xùn)練算法 ..........................................................121
3.3.5卷積網(wǎng)提取特征的可視化 .............................................129
3.4 CNNs的變體 ..............................................................131
3.4.1關(guān)于卷積核的變體 .....................................................132
3.4.2關(guān)于卷積通道的變體 ..................................................136
3.4.3關(guān)于卷積層連接的變體 .................................................141
3.5卷積網(wǎng)絡(luò)在自動駕駛中的應(yīng)用 .............................................145
3.6 卷積網(wǎng)絡(luò)發(fā)展史 .........................................................146
參考文獻(xiàn) ................................................................148

第 4章反饋神經(jīng)網(wǎng)絡(luò) ........................................................151
4.1引言 ....................................................................151
4.2反饋神經(jīng)網(wǎng)絡(luò) ............................................................153
4.2.1統(tǒng)計語言模型 .......................................................154
4.2.2 RNNs的網(wǎng)絡(luò)結(jié)構(gòu) ........................................................158
4.2.3 RNNs的訓(xùn)練算法：BPTT ...............................................162
4.2.4 RNNs的誤差沿時間軸傳播公式 .........................................170
4.2.5 RNNs的變體 ..........................................................172
4.3長短期記憶單元反饋神經(jīng)網(wǎng)絡(luò) ............................................175
4.3.1早期解決長時間序列學(xué)習(xí)難題的樸素方法 .................................176
4.3.2 LSTM網(wǎng)絡(luò)結(jié)構(gòu) .........................................................177
4.3.3 LSTM的前向計算 ....................................................179
4.3.4 LSTM的反向計算 .....................................................181
4.3.5 LSTM的權(quán)值更新 .....................................................184
4.3.6理解 LSTM網(wǎng)絡(luò) ........................................................186
4.4時間序列處理中的幾種重要機制 ..........................................188
4.4.1處理變長序列的編碼器-解碼器模型 ..................................189
4.4.2注意力機制 ............................................................192
4.4.3序列自動對齊的 CTC技術(shù) .............................................195
4.4.4小結(jié) .............................................................215
4.5深度反饋網(wǎng)絡(luò)在時間序列處理中的應(yīng)用 ....................................216
4.5.1 Google神經(jīng)機器翻譯系統(tǒng) ...............................................216
4.5.2深度語音識別系統(tǒng) ..................................................219
4.5.3用 LSTM編程和寫作 ..................................................223
4.6 反饋神經(jīng)網(wǎng)絡(luò)發(fā)展現(xiàn)狀 ....................................................229
參考文獻(xiàn) ...................................................................229

第 5章深度強化學(xué)習(xí) .........................................................234
5.1引言 ....................................................................234
5.2馬爾可夫決策過程 .....................................................238
5.2.1馬爾可夫過程 .......................................................239
5.2.2馬爾可夫獎勵過程 .....................................................243
5.2.3馬爾可夫決策過程 .....................................................246
5.2.4廣義策略迭代 ..........................................................255
5.2.5小結(jié) .................................................................257
5.3強化學(xué)習(xí)算法 ............................................................257
5.3.1動態(tài)規(guī)劃算法 ........................................................258
5.3.2蒙特卡羅算法 ..........................................................263
5.3.3時間差分學(xué)習(xí)算法 ......................................................274
5.3.4小結(jié) ..................................................................297
5.4深度強化學(xué)習(xí)算法 ........................................................299
5.4.1基于深度網(wǎng)絡(luò)的狀態(tài)價值和動作價值函數(shù)近似 ...........................301
5.4.2基于深度網(wǎng)絡(luò)的策略梯度法 ..........................................307
5.5深度強化學(xué)習(xí)的應(yīng)用 .....................................................317
5.5.1圍棋 AlphaGo ........................................................317
5.5.2從 AlphaGo到 AlphaGo Zero............................................332
5.5.3基于像素的乒乓球游戲 .................................................341
5.6深度強化學(xué)習(xí)發(fā)展現(xiàn)狀 ................................................344
5.6.1強化學(xué)習(xí)起源與發(fā)展現(xiàn)狀 ................................................344
5.6.2深度強化學(xué)習(xí)在棋類中的應(yīng)用 ............................................345
5.6.3深度強化學(xué)習(xí)技術(shù)在自主智能體中的應(yīng)用 .................................347
參考文獻(xiàn) ..................................................................350
后記 ........................................................................353