%PDF-1.3 1 0 obj << /Kids [ 4 0 R 5 0 R 6 0 R 7 0 R 8 0 R 9 0 R 10 0 R 11 0 R 12 0 R 13 0 R 14 0 R ] /Type /Pages /Count 11 >> endobj 2 0 obj << /Subject (Neural Information Processing Systems http\072\057\057nips\056cc\057) /Publisher (Curran Associates\054 Inc\056) /Language (en\055US) /Created (2018) /EventType (Poster) /Description-Abstract (We present SplineNets\054 a practical and novel approach for using conditioning in convolutional neural networks \050CNNs\051\056 SplineNets are continuous generalizations of neural decision graphs\054 and they can dramatically reduce runtime complexity and computation costs of CNNs\054 while maintaining or even increasing accuracy\056 Functions of SplineNets are both dynamic \050i\056e\056\054 conditioned on the input\051 and hierarchical \050i\056e\056\054conditioned on the computational path\051\056 SplineNets employ a unified loss function with a desired level of smoothness over both the network and decision parameters\054 while allowing for sparse activation of a subset of nodes for individual samples\056 In particular\054 we embed infinitely many function weights \050e\056g\056 filters\051 on smooth\054 low dimensional manifolds parameterized by compact B\055splines\054 which are indexed by a position parameter\056 Instead of sampling from a categorical distribution to pick a branch\054 samples choose a continuous position to pick a function weight\056 We further show that by maximizing the mutual information between spline positions and class labels\054 the network can be optimally utilized and specialized for classification tasks\056 Experiments show that our approach can significantly increase the accuracy of ResNets with negligible cost in speed\054 matching the precision of a 110 level ResNet with a 32 level SplineNet\056) /Producer (PyPDF2) /Title (SplineNets\072 Continuous Neural Decision Graphs) /Date (2018) /ModDate (D\07220190218213719\05508\04700\047) /Published (2018) /Type (Conference Proceedings) /firstpage (1994) /Book (Advances in Neural Information Processing Systems 31) /Description (Paper accepted and presented at the Neural Information Processing Systems Conference \050http\072\057\057nips\056cc\057\051) /Editors (S\056 Bengio and H\056 Wallach and H\056 Larochelle and K\056 Grauman and N\056 Cesa\055Bianchi and R\056 Garnett) /Author (Cem Keskin\054 Shahram Izadi) /lastpage (2004) >> endobj 3 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 4 0 obj << /Contents 15 0 R /Parent 1 0 R /Resources 16 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 34 0 R 35 0 R 36 0 R 37 0 R 38 0 R 39 0 R 40 0 R ] /Type /Page >> endobj 5 0 obj << /Contents 41 0 R /Parent 1 0 R /Resources 42 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 47 0 R 48 0 R 49 0 R 50 0 R 51 0 R 52 0 R 53 0 R 54 0 R 55 0 R 56 0 R 57 0 R 58 0 R 59 0 R 60 0 R 61 0 R ] /Type /Page >> endobj 6 0 obj << /Contents 62 0 R /Parent 1 0 R /Resources 63 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 97 0 R 98 0 R ] /Type /Page >> endobj 7 0 obj << /Contents 99 0 R /Parent 1 0 R /Resources 100 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 103 0 R 104 0 R ] /Type /Page >> endobj 8 0 obj << /Contents 105 0 R /Parent 1 0 R /Resources 106 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 111 0 R 112 0 R ] /Type /Page >> endobj 9 0 obj << /Contents 113 0 R /Parent 1 0 R /Resources 114 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 121 0 R ] /Type /Page >> endobj 10 0 obj << /Contents 122 0 R /Parent 1 0 R /Resources 123 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 131 0 R 132 0 R 133 0 R 134 0 R 135 0 R 136 0 R ] /Type /Page >> endobj 11 0 obj << /Contents 137 0 R /Parent 1 0 R /Resources 138 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 145 0 R 146 0 R 147 0 R ] /Type /Page >> endobj 12 0 obj << /Contents 148 0 R /Parent 1 0 R /Resources 149 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 154 0 R 155 0 R 156 0 R 157 0 R 158 0 R 159 0 R 160 0 R 161 0 R 162 0 R ] /Type /Page >> endobj 13 0 obj << /Contents 163 0 R /Parent 1 0 R /Type /Page /Resources 164 0 R /MediaBox [ 0 0 612 792 ] >> endobj 14 0 obj << /Contents 165 0 R /Parent 1 0 R /Type /Page /Resources 166 0 R /MediaBox [ 0 0 612 792 ] >> endobj 15 0 obj << /Length 3114 /Filter /FlateDecode >> stream xڽYKsFWU0gN;vZ'*[% #kb߯ @A aޏn&|nB|M l