%PDF-1.4 % 1 0 obj << /S /GoTo /D (section*.2) >> endobj 4 0 obj (Contents) endobj 5 0 obj << /S /GoTo /D (chapter.1) >> endobj 8 0 obj (1 Introduction) endobj 9 0 obj << /S /GoTo /D (section.1.1) >> endobj 12 0 obj (1.1 Introduction) endobj 13 0 obj << /S /GoTo /D (section.1.2) >> endobj 16 0 obj (1.2 What is the transforming autoencoder) endobj 17 0 obj << /S /GoTo /D (subsection.1.2.1) >> endobj 20 0 obj (1.2.1 The autoencoder) endobj 21 0 obj << /S /GoTo /D (subsection.1.2.2) >> endobj 24 0 obj (1.2.2 The transforming autoencoder) endobj 25 0 obj << /S /GoTo /D (section.1.3) >> endobj 28 0 obj (1.3 Relationship to other works) endobj 29 0 obj << /S /GoTo /D (chapter.2) >> endobj 32 0 obj (2 The transforming autoencoder) endobj 33 0 obj << /S /GoTo /D (section.2.1) >> endobj 36 0 obj (2.1 The architecture of transforming autoencoder) endobj 37 0 obj << /S /GoTo /D (subsection.2.1.1) >> endobj 40 0 obj (2.1.1 A high level overview) endobj 41 0 obj << /S /GoTo /D (subsection.2.1.2) >> endobj 44 0 obj (2.1.2 Details of the transforming autoencoder) endobj 45 0 obj << /S /GoTo /D (section.2.2) >> endobj 48 0 obj (2.2 Training the transforming autoencoder) endobj 49 0 obj << /S /GoTo /D (subsection.2.2.1) >> endobj 52 0 obj (2.2.1 How to train) endobj 53 0 obj << /S /GoTo /D (subsection.2.2.2) >> endobj 56 0 obj (2.2.2 The transforming autoencoder on translations of MNIST) endobj 57 0 obj << /S /GoTo /D (subsection.2.2.3) >> endobj 60 0 obj (2.2.3 The transforming autoencoder on more complex transformations) endobj 61 0 obj << /S /GoTo /D (section.2.3) >> endobj 64 0 obj (2.3 How does the transforming autoencoder work) endobj 65 0 obj << /S /GoTo /D (chapter.3) >> endobj 68 0 obj (3 Using the transforming autoencoder for classification) endobj 69 0 obj << /S /GoTo /D (section.3.1) >> endobj 72 0 obj (3.1 Classification for MNIST digits) endobj 73 0 obj << /S /GoTo /D (section.3.2) >> endobj 76 0 obj (3.2 Applying the transforming autoencoder to MNIST classification) endobj 77 0 obj << /S /GoTo /D (subsection.3.2.1) >> endobj 80 0 obj (3.2.1 Baseline results) endobj 81 0 obj << /S /GoTo /D (subsection.3.2.2) >> endobj 84 0 obj (3.2.2 Classification using pairwise differences) endobj 85 0 obj << /S /GoTo /D (subsection.3.2.3) >> endobj 88 0 obj (3.2.3 An overview of how pairwise differences work) endobj 89 0 obj << /S /GoTo /D (subsection.3.2.4) >> endobj 92 0 obj (3.2.4 Mathematical details) endobj 93 0 obj << /S /GoTo /D (section.3.3) >> endobj 96 0 obj (3.3 Generalizing to translated test data) endobj 97 0 obj << /S /GoTo /D (subsection.3.3.1) >> endobj 100 0 obj (3.3.1 Description of this task) endobj 101 0 obj << /S /GoTo /D (subsection.3.3.2) >> endobj 104 0 obj (3.3.2 Results on this task) endobj 105 0 obj << /S /GoTo /D (chapter.4) >> endobj 108 0 obj (4 Tricks and future works) endobj 109 0 obj << /S /GoTo /D (section.4.1) >> endobj 112 0 obj (4.1 Improving performance by adding layers) endobj 113 0 obj << /S /GoTo /D (section.4.2) >> endobj 116 0 obj (4.2 Regularized transforming autoencoder) endobj 117 0 obj << /S /GoTo /D (section.4.3) >> endobj 120 0 obj (4.3 Deeper transforming autoencoder) endobj 121 0 obj << /S /GoTo /D (section.4.4) >> endobj 124 0 obj (4.4 Transforming autoencoder with local fields) endobj 125 0 obj << /S /GoTo /D (chapter.5) >> endobj 128 0 obj (5 Conclusions) endobj 129 0 obj << /S /GoTo /D (appendix.A) >> endobj 132 0 obj (A Neural network and training details) endobj 133 0 obj << /S /GoTo /D (section.A.1) >> endobj 136 0 obj (A.1 Neural network) endobj 137 0 obj << /S /GoTo /D (subsection.A.1.1) >> endobj 140 0 obj (A.1.1 Back-propagation using stochastic gradient descent) endobj 141 0 obj << /S /GoTo /D (section.A.2) >> endobj 144 0 obj (A.2 Generating translated/deformed data) endobj 145 0 obj << /S /GoTo /D (section.A.3) >> endobj 148 0 obj (A.3 Miscellaneous techniques used in this thesis) endobj 149 0 obj << /S /GoTo /D (subsection.A.3.1) >> endobj 152 0 obj (A.3.1 Weight decay) endobj 153 0 obj << /S /GoTo /D (subsection.A.3.2) >> endobj 156 0 obj (A.3.2 Curriculum learning) endobj 157 0 obj << /S /GoTo /D (subsection.A.3.3) >> endobj 160 0 obj (A.3.3 Learning schedule) endobj 161 0 obj << /S /GoTo /D (section.A.4) >> endobj 164 0 obj (A.4 Terminologies) endobj 165 0 obj << /S /GoTo /D (appendix.B) >> endobj 168 0 obj (B Detailed classification results) endobj 169 0 obj << /S /GoTo /D (section.B.1) >> endobj 172 0 obj (B.1 Table of results) endobj 173 0 obj << /S /GoTo /D (section.B.2) >> endobj 176 0 obj (B.2 Error cases) endobj 177 0 obj << /S /GoTo /D (section.B.3) >> endobj 180 0 obj (B.3 Error cases when generalizing to translated test data) endobj 181 0 obj << /S /GoTo /D (appendix.C) >> endobj 184 0 obj (C Reproducing results in this thesis) endobj 185 0 obj << /S /GoTo /D (section.C.1) >> endobj 188 0 obj (C.1 Prerequisites) endobj 189 0 obj << /S /GoTo /D (section.C.2) >> endobj 192 0 obj (C.2 Obtaining the code and weights) endobj 193 0 obj << /S /GoTo /D (section*.4) >> endobj 196 0 obj (References) endobj 197 0 obj << /S /GoTo /D [198 0 R /Fit ] >> endobj 203 0 obj << /Length 426 /Filter /FlateDecode >> stream xڅn0E .DQˤ]..i"w(Fj(>Ν{G #wzl"BT4;MEDAҤX?a*0zmCJ %Pe5Os3m{.9<ŋ $-EVε@n{!1]tY^QK,13PVc"W&p5+*[^L()3Lև\TY@ٸKOG'}=#JezY~?BMnx60q%pxFIQKJN3. w!,.t ge >!U>a,uzr\ui#7 //%0'wԎ`6fC- endstream endobj 198 0 obj << /Type /Page /Contents 203 0 R /Resources 202 0 R /MediaBox [0 0 595.276 841.89] /Parent 210 0 R /Annots [ 199 0 R 200 0 R 201 0 R ] >> endobj 199 0 obj << /Type /Annot /Border[0 0 0]/H/I/C[0 1 1] /Rect [270.075 415.914 346.716 433.21] /Subtype/Link/A<> >> endobj 200 0 obj << /Type /Annot /Border[0 0 0]/H/I/C[0 1 1] /Rect [204.876 388.041 411.916 402.786] /Subtype/Link/A<> >> endobj 201 0 obj << /Type /Annot /Border[0 0 0]/H/I/C[0 1 1] /Rect [241.42 359.61 375.372 374.355] /Subtype/Link/A<> >> endobj 204 0 obj << /D [198 0 R /XYZ 101.884 768.055 null] >> endobj 205 0 obj << /D [198 0 R /XYZ 102.884 699.659 null] >> endobj 202 0 obj << /Font << /F17 206 0 R /F19 207 0 R /F18 208 0 R /F20 209 0 R >> /ProcSet [ /PDF /Text ] >> endobj 215 0 obj << /Length 949 /Filter /FlateDecode >> stream x}UKo6WV ]l阢]7mDKl(M_z^p