%PDF-1.5 % 1 0 obj << /S /GoTo /D (section.1) >> endobj 4 0 obj (Introduction) endobj 5 0 obj << /S /GoTo /D (section.2) >> endobj 8 0 obj (Pruning channels and corresponding feature maps) endobj 9 0 obj << /S /GoTo /D (subsection.2.1) >> endobj 12 0 obj (Channel pruning for networks without shortcut connections) endobj 13 0 obj << /S /GoTo /D (subsection.2.2) >> endobj 16 0 obj (Channel pruning for residual networks) endobj 17 0 obj << /S /GoTo /D (section.3) >> endobj 20 0 obj (Channel pruning strategy) endobj 21 0 obj << /S /GoTo /D (subsection.3.1) >> endobj 24 0 obj (Mean gradient criterion for pruning) endobj 25 0 obj << /S /GoTo /D (subsection.3.2) >> endobj 28 0 obj (Hierarchical global pruning strategy) endobj 29 0 obj << /S /GoTo /D (subsection.3.3) >> endobj 32 0 obj (Fine-tuning of the pruned networks) endobj 33 0 obj << /S /GoTo /D (section.4) >> endobj 36 0 obj (Experiments) endobj 37 0 obj << /S /GoTo /D (subsection.4.1) >> endobj 40 0 obj (Effect of mean gradient on network performance) endobj 41 0 obj << /S /GoTo /D (subsection.4.2) >> endobj 44 0 obj (Comparison with global pruning strategy) endobj 45 0 obj << /S /GoTo /D (subsection.4.3) >> endobj 48 0 obj (Comparison with other pruning criteria) endobj 49 0 obj << /S /GoTo /D (subsection.4.4) >> endobj 52 0 obj (Pruning VGG-16 on CIFAR-10) endobj 53 0 obj << /S /GoTo /D (subsection.4.5) >> endobj 56 0 obj (Pruning ResNet-110 on CIFAR-10) endobj 57 0 obj << /S /GoTo /D (section.5) >> endobj 60 0 obj (Conclusion) endobj 61 0 obj << /S /GoTo /D [62 0 R /Fit] >> endobj 78 0 obj << /Length 2200 /Filter /FlateDecode >> stream xڭr8e@TLzw0 KCˇvAh~ۀ?]Q*2e"ǀ8Y$:, >|3Mc:IxۍMlq?՚ۆ0!8LQfmGxyn+ۙn~ n5O,i%tVP';v3vGrootEY"0OUYooZ+?.`K?VZ#q jJcM WbTq Ҥ#Z8t"fS6E } ,+rv]ܕlh|]_ϗNcX*m%ջ]88xH*M" 7,(׀Eqì)۳ uOr,.>RiĴ$3HT$@]fz@CC;WػOzrB[;8s=BCY,zߵCZNG)ky6^`tE*atƞvewv:#1z?U7ԑaV:E