%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 15 0 R ] /Type /Pages /Count 12 >> 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 (2019) /EventType (Poster) /Description-Abstract (Sum\055product networks \050SPNs\051 are flexible density estimators and have received significant attention due to their attractive inference properties\056 While parameter learning in SPNs is well developed\054 structure learning leaves something to be desired\072 Even though there is a plethora of SPN structure learners\054 most of them are somewhat ad\055hoc and based on intuition rather than a clear learning principle\056 In this paper\054 we introduce a well\055principled Bayesian framework for SPN structure learning\056 First\054 we decompose the problem into i\051 laying out a computational graph\054 and ii\051 learning the so\055called scope function over the graph\056 The first is rather unproblematic and akin to neural network architecture validation\056 The second represents the effective structure of the SPN and needs to respect the usual structural constraints in SPN\054 i\056e\056 completeness and decomposability\056 While representing and learning the scope function is somewhat involved in general\054 in this paper\054 we propose a natural parametrisation for an important and widely used special case of SPNs\056 These structural parameters are incorporated into a Bayesian model\054 such that simultaneous structure and parameter learning is cast into monolithic Bayesian posterior inference\056 In various experiments\054 our Bayesian SPNs often improve test likelihoods over greedy SPN learners\056 Further\054 since the Bayesian framework protects against overfitting\054 we can evaluate hyper\055parameters directly on the Bayesian model score\054 waiving the need for a separate validation set\054 which is especially beneficial in low data regimes\056 Bayesian SPNs can be applied to heterogeneous domains and can easily be extended to nonparametric formulations\056 Moreover\054 our Bayesian approach is the first\054 which consistently and robustly learns SPN structures under missing data\056) /Producer (PyPDF2) /Title (Bayesian Learning of Sum\055Product Networks) /Date (2019) /ModDate (D\07220200213004509\05508\04700\047) /Published (2019) /Type (Conference Proceedings) /firstpage (6347) /Book (Advances in Neural Information Processing Systems 32) /Description (Paper accepted and presented at the Neural Information Processing Systems Conference \050http\072\057\057nips\056cc\057\051) /Editors (H\056 Wallach and H\056 Larochelle and A\056 Beygelzimer and F\056 d\047Alch\351\055Buc and E\056 Fox and R\056 Garnett) /Author (Martin Trapp\054 Robert Peharz\054 Hong Ge\054 Franz Pernkopf\054 Zoubin Ghahramani) /lastpage (6358) >> endobj 3 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 4 0 obj << /Contents 16 0 R /Parent 1 0 R /Resources 17 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 62 0 R 63 0 R ] /Type /Page >> endobj 5 0 obj << /Contents 64 0 R /Parent 1 0 R /Resources 65 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 102 0 R 103 0 R 104 0 R 105 0 R 106 0 R 107 0 R 108 0 R 109 0 R 110 0 R 111 0 R 112 0 R 113 0 R 114 0 R 115 0 R 116 0 R 117 0 R 118 0 R 119 0 R 120 0 R 121 0 R ] /Type /Page >> endobj 6 0 obj << /Contents 122 0 R /Parent 1 0 R /Resources 123 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 160 0 R 161 0 R 162 0 R 163 0 R 164 0 R 165 0 R 166 0 R 167 0 R 168 0 R 169 0 R 170 0 R 171 0 R 172 0 R 173 0 R 174 0 R 175 0 R 176 0 R 177 0 R 178 0 R 179 0 R 180 0 R 181 0 R 182 0 R ] /Type /Page >> endobj 7 0 obj << /Contents 183 0 R /Parent 1 0 R /Resources 184 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 189 0 R 190 0 R 191 0 R 192 0 R 193 0 R 194 0 R 195 0 R 196 0 R 197 0 R 198 0 R 199 0 R ] /Type /Page >> endobj 8 0 obj << /Contents 200 0 R /Parent 1 0 R /Resources 201 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 210 0 R 211 0 R 212 0 R 213 0 R 214 0 R 215 0 R 216 0 R 217 0 R ] /Type /Page >> endobj 9 0 obj << /Contents 218 0 R /Parent 1 0 R /Resources 219 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 229 0 R 230 0 R 231 0 R 232 0 R 233 0 R 234 0 R 235 0 R 236 0 R 237 0 R 238 0 R 239 0 R ] /Type /Page >> endobj 10 0 obj << /Contents 240 0 R /Parent 1 0 R /Resources 241 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 254 0 R 255 0 R 256 0 R 257 0 R 258 0 R 259 0 R 260 0 R 261 0 R 262 0 R 263 0 R 264 0 R 265 0 R ] /Type /Page >> endobj 11 0 obj << /Contents 266 0 R /Parent 1 0 R /Resources 267 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 272 0 R 273 0 R 274 0 R 275 0 R 276 0 R 277 0 R 278 0 R 279 0 R 280 0 R 281 0 R 282 0 R 283 0 R 284 0 R 285 0 R 286 0 R 287 0 R 288 0 R 289 0 R 290 0 R 291 0 R 292 0 R ] /Type /Page >> endobj 12 0 obj << /Contents 293 0 R /Parent 1 0 R /Resources 294 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 299 0 R 300 0 R 301 0 R ] /Type /Page >> endobj 13 0 obj << /Contents 302 0 R /Parent 1 0 R /Type /Page /Resources 303 0 R /MediaBox [ 0 0 612 792 ] >> endobj 14 0 obj << /Contents 304 0 R /Parent 1 0 R /Type /Page /Resources 305 0 R /MediaBox [ 0 0 612 792 ] >> endobj 15 0 obj << /Contents 306 0 R /Parent 1 0 R /Type /Page /Resources 307 0 R /MediaBox [ 0 0 612 792 ] >> endobj 16 0 obj << /Length 3581 /Filter /FlateDecode >> stream xڥZMFБ(qڽp3]vXRV!)˿~ Te3 x@Wûdw JTRwD4e2qYέݟ}W*JiL'E".S߭~\UFIWWukFM<3}Q;o&膯ġ~!jq?=$Y%q YG/}:>q*1ʴU)bPuC-߯suD}27-FLS'$$\tll|LB| vc>Nia}vAlj.B1]6Վ']3:.VUXU)ﻪ-TmOO7%C6;df o{~OþwձjmJm]!l[aTi]%'DiYu}=~SnM{h_YG0NCb=Eɷo( `@n8ط4:>vR$??Sۏ sY T d=^7UT 0LSZlc\mDT~07P\-mwo\O'ׇ7iIngWK&<}!/pPu fR<8MtlA?A`6(69Iϟ~xN9$!Z?O߹f