%PDF-1.3 1 0 obj << /Kids [ 3 0 R 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 ] /Type /Pages /Count 10 >> endobj 2 0 obj << /Title (Neural Module Networks) /Producer (PyPDF2) /Author (Jacob Andreas\054 Marcus Rohrbach\054 Trevor Darrell\054 Dan Klein) /Subject (2016 IEEE Conference on Computer Vision and Pattern Recognition) >> endobj 3 0 obj << /Parent 1 0 R /Rotate 0 /Contents 14 0 R /Resources << /XObject << /x8 15 0 R /R49 18 0 R /R54 21 0 R /R55 22 0 R /R56 23 0 R /R57 24 0 R /R50 25 0 R /R51 26 0 R /R52 27 0 R /R53 28 0 R /R46 29 0 R /x12 30 0 R /x6 33 0 R /x10 36 0 R /R62 39 0 R >> /ExtGState << /s9 66 0 R /R48 19 0 R /R45 42 0 R /s5 69 0 R /s7 72 0 R /R40 43 0 R /s11 75 0 R /a0 << /CA 1 /ca 1 >> /R34 40 0 R /R35 41 0 R /R23 78 0 R >> /Font << /R59 58 0 R /R42 51 0 R /F2 79 0 R /R28 80 0 R /F1 84 0 R /R37 44 0 R /R32 85 0 R /R24 89 0 R /R30 92 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 65 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 99 0 R 100 0 R 101 0 R 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 ] /Type /Page >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 112 0 R /Resources << /ExtGState << /R23 78 0 R >> /Font << /R86 113 0 R /F2 118 0 R /F1 119 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 120 0 R 121 0 R 122 0 R 123 0 R 124 0 R 125 0 R 126 0 R 127 0 R 128 0 R 129 0 R 130 0 R 131 0 R 132 0 R 133 0 R 134 0 R 135 0 R 136 0 R ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Contents 137 0 R /Resources << /XObject << /R130 138 0 R /R131 139 0 R /R129 140 0 R >> /ExtGState << /R23 78 0 R /R35 41 0 R /R34 40 0 R /R45 42 0 R /R48 19 0 R >> /Font << /F2 141 0 R /R119 142 0 R /R59 58 0 R /R117 146 0 R /R115 150 0 R /R127 153 0 R /R125 156 0 R /F1 160 0 R /R123 161 0 R /R121 164 0 R /R37 44 0 R /R24 89 0 R /R32 85 0 R /R86 113 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 65 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 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 183 0 R 184 0 R 185 0 R 186 0 R 187 0 R 188 0 R ] /Type /Page >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Contents 189 0 R /Resources << /XObject << /R145 190 0 R /R162 191 0 R /R147 192 0 R /R146 193 0 R /R156 194 0 R /R157 195 0 R /R154 196 0 R /R160 202 0 R /R152 198 0 R /R153 199 0 R /R150 200 0 R /R164 203 0 R /R165 204 0 R /R149 197 0 R /R148 207 0 R /R158 208 0 R /R159 209 0 R /R161 206 0 R /R151 201 0 R /R163 205 0 R /R155 210 0 R >> /ExtGState << /R23 78 0 R /R35 41 0 R /R34 40 0 R /R45 42 0 R /R48 19 0 R >> /Font << /F2 211 0 R /R119 142 0 R /R86 113 0 R /R59 58 0 R /R42 51 0 R /R117 146 0 R /R127 153 0 R /R125 156 0 R /F1 212 0 R /R123 161 0 R /R166 213 0 R /R37 44 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 65 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 216 0 R 217 0 R 218 0 R 219 0 R 220 0 R 221 0 R ] /Type /Page >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Contents 222 0 R /Resources << /XObject << /R192 223 0 R /R193 224 0 R /R190 225 0 R /R191 226 0 R /R196 227 0 R /R197 228 0 R /R194 229 0 R /R195 230 0 R /R189 231 0 R /R188 232 0 R /R198 233 0 R /R183 236 0 R /R182 237 0 R /R185 238 0 R /R184 242 0 R /R179 241 0 R /R180 240 0 R /R187 234 0 R /R181 239 0 R /R186 235 0 R >> /ExtGState << /R23 78 0 R /R35 41 0 R /R34 40 0 R /R45 42 0 R /R48 19 0 R >> /Font << /F2 243 0 R /R42 51 0 R /R86 113 0 R /R59 58 0 R /R37 44 0 R /F1 244 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 65 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 245 0 R 246 0 R 247 0 R 248 0 R ] /Type /Page >> endobj 8 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 249 0 R /Resources << /ExtGState << /R23 78 0 R >> /Font << /R86 113 0 R /F2 250 0 R /F1 251 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 252 0 R 253 0 R 254 0 R 255 0 R 256 0 R 257 0 R 258 0 R 259 0 R 260 0 R ] >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 261 0 R /Resources << /ExtGState << /R23 78 0 R >> /Font << /R229 262 0 R /R86 113 0 R /F2 266 0 R /F1 267 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 268 0 R 269 0 R 270 0 R 271 0 R 272 0 R 273 0 R 274 0 R 275 0 R 276 0 R 277 0 R 278 0 R ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 279 0 R /Resources << /XObject << /R239 280 0 R /R238 281 0 R /R237 282 0 R /R248 283 0 R /R246 285 0 R /R244 286 0 R /R245 287 0 R /R242 288 0 R /R243 290 0 R /R240 291 0 R >> /ExtGState << /R23 78 0 R >> /Font << /F2 292 0 R /F1 293 0 R /R32 85 0 R /R86 113 0 R /R26 95 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] /ColorSpace << /R247 284 0 R /R241 289 0 R >> >> /MediaBox [ 0 0 612 792 ] /Annots [ 294 0 R 295 0 R ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 296 0 R /Resources << /ExtGState << /R23 78 0 R >> /Font << /F2 297 0 R /F1 298 0 R /R32 85 0 R /R24 89 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 299 0 R 300 0 R 301 0 R 302 0 R 303 0 R 304 0 R 305 0 R 306 0 R 307 0 R 308 0 R 309 0 R 310 0 R 311 0 R 312 0 R 313 0 R 314 0 R 315 0 R 316 0 R 317 0 R 318 0 R 319 0 R 320 0 R 321 0 R 322 0 R 323 0 R 324 0 R 325 0 R 326 0 R 327 0 R 328 0 R 329 0 R 330 0 R 331 0 R 332 0 R 333 0 R ] >> endobj 12 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 334 0 R /Resources << /ExtGState << /R23 78 0 R >> /Font << /F2 335 0 R /F1 336 0 R /R32 85 0 R /R26 95 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 337 0 R 338 0 R 339 0 R 340 0 R 341 0 R 342 0 R 343 0 R 344 0 R 345 0 R 346 0 R 347 0 R 348 0 R 349 0 R 350 0 R 351 0 R 352 0 R 353 0 R 354 0 R 355 0 R 356 0 R ] >> endobj 13 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 14 0 obj << /Length 16641 >> stream q q q 0.10000 0 0 0.10000 0 0 cm /R23 gs 0 g q 10 0 0 10 0 0 cm BT /R24 14.34620 Tf 1 0 0 1 219.98500 675.06700 Tm [ (Neural) -250.00800 (Module) -249.99100 (Netw) 9.99455 (orks) ] TJ /R26 11.95520 Tf -85.90120 -37.85820 Td [ (Jacob) -250.01300 (Andreas) -1500 (Marcus) -250.01200 (Rohrbach) -1499.99000 (T) 34.98980 (re) 25.00590 (v) 19.98430 (or) -250.01400 (Darrell) -1499.99000 (Dan) -250.00600 (Klein) ] TJ 81.90510 -13.94690 Td [ (Uni) 24.99570 (v) 14.98510 (ersity) -249.98900 (of) -250.01400 (California\054) -249.99300 (Berk) 9.99010 (ele) 14.99130 (y) ] TJ /R28 8.96640 Tf -38.65120 -13.94800 Td [ (f) -0.89990 ] TJ /R30 8.96640 Tf 4.60820 0 Td (jda\054rohrbach\054trevor\054klein) Tj /R28 8.96640 Tf 134.49500 0 Td [ (g) -0.90126 ] TJ /R30 8.96640 Tf 4.60781 0 Td (\100eecs\056berkeley\056edu) Tj /R24 11.95520 Tf -175.05400 -41.04610 Td (Abstract) Tj /R32 9.96260 Tf -83.92810 -24.25780 Td [ (V) 73.99130 (isual) -484.00600 (question) -484.00100 (answering) -483.98900 (is) -485.00400 (fundamentally) -483.98300 (composi\055) ] TJ -11.95470 -11.95630 Td [ (tional) -309.99000 (in) -308.99500 (natur) 37.00240 (e\227a) -310.01400 (question) -309.01100 (lik) 10.00810 (e) ] TJ /R26 9.96260 Tf 133.95700 0 Td [ (where) -309.99200 (is) -308.99500 (the) -310.01900 (dog\077) ] TJ /R32 9.96260 Tf 76.64880 0 Td [ (shar) 36.99630 (es) ] TJ -210.60600 -11.95510 Td [ (substructur) 36.98950 (e) -266 (with) -266 (questions) -265.01200 (lik) 10.00810 (e) ] TJ /R26 9.96260 Tf 129.12600 0 Td [ (what) -265.98400 (color) -265.98500 (is) -265.00200 (the) -265.98700 (dog\077) ] TJ /R32 9.96260 Tf 92.18010 0 Td (and) Tj /R26 9.96260 Tf -221.30600 -11.95470 Td [ (where) -228.00500 (is) -227.98800 (the) -227.99300 (cat\077) ] TJ /R32 9.96260 Tf 69.02380 0 Td [ (This) -228.00700 (paper) -227.98400 (seeks) -228.00200 (to) -227.98900 (simultaneousl) 1.00473 (y) -228.00600 (e) 19.99180 (xploit) ] TJ -69.02380 -11.95510 Td [ (the) -195.99900 (r) 37.01960 (epr) 36.98100 (esentational) -195.99400 (capacity) -195.99100 (of) -195.99400 (deep) -196.00500 (networks) -196.01100 (and) -197.01100 (the) -196 (com\055) ] TJ 11.95510 TL T* [ (positional) -330.98300 (linguistic) -330.01000 (structur) 37.01220 (e) -331.00800 (of) -330.99100 (questions\056) -551.98100 (W) 91.98710 (e) -331.00800 (describe) -330.98600 (a) ] TJ 11.95590 TL T* [ (pr) 44.98390 (ocedur) 36.98650 (e) -224.98600 (for) -225.00300 (constructing) -225.01000 (and) -225.00700 (learning) ] TJ /R26 9.96260 Tf 161.48900 0 Td [ (neural) -225.01600 (module) -224.99200 (net\055) ] TJ -161.48900 -11.95510 Td [ (w) 10.00140 (orks) ] TJ /R32 9.96260 Tf 24.24880 0 Td [ (\054) -287.00800 (whic) 14.99870 (h) -279.98800 (compose) -280.01700 (collections) -279.99500 (of) -280.01900 (jointly\055tr) 14.98280 (ained) -280.01500 (neur) 14.99010 (al) ] TJ -24.24880 -11.95510 Td [ (\223modules\224) -223.98500 (into) -224.01700 (deep) -224 (networks) -224.00700 (for) -223.98300 (quest) 1 (ion) -223.98500 (answering) 14.98770 (\056) -301.98600 (Our) ] TJ 11.95510 TL T* [ (appr) 44.99370 (oac) 14.98340 (h) -406.98600 (decomposes) -407.00800 (questions) -406.00800 (into) -407.00600 (their) -407.00800 (linguistic) -407.01800 (sub\055) ] TJ T* [ (structur) 37.01280 (es\054) -261.01100 (and) -259.98100 (uses) -259.00300 (these) -259.00600 (structur) 37.01220 (es) -259.01400 (to) -259.98400 (dynamical) 1.00964 (ly) -259.99300 (instan\055) ] TJ T* [ (tiate) -278.99000 (modular) -279.99600 (networks) -279.01700 (\050with) -280.00200 (r) 37.01960 (eusable) -278.99300 (components) -279.02000 (for) -280.01500 (r) 37.01830 (ec\055) ] TJ 11.95590 TL T* [ (o) 10.00320 (gnizing) -281.99100 (do) 10.00810 (gs\054) -290.01700 (classifying) -281.98400 (color) 10.01670 (s\054) -289.98200 (etc\056\051\056) -406.99100 (The) -282.01700 (r) 37.01830 (esulting) -281.99200 (com\055) ] TJ 11.95510 TL T* [ (pound) -379.01000 (networks) -377.98100 (ar) 36.98520 (e) -379 (jointly) -379 (tr) 14.99140 (ained\056) -695.98100 (W) 91.98710 (e) -377.98100 (e) 15.01220 (valuate) -379.01500 (our) -379.01800 (ap\055) ] TJ T* [ (pr) 44.98390 (oac) 14.98340 (h) -279.00700 (on) -279.00200 (two) -278.98700 (c) 15.01220 (hallenging) -279.00400 (datasets) -279.01500 (for) -280.01500 (visual) -278.98500 (question) -279.01500 (an\055) ] TJ T* [ (swering) 9.99772 (\054) -261.99300 (ac) 15.01770 (hie) 14.98520 (ving) -259.98400 (state\055of\055the\055art) -259.00500 (r) 37.01960 (esults) -260.00800 (on) -259.00600 (both) -260.01300 (the) -259.98800 (VQA) ] TJ T* [ (natur) 15.00610 (al) -302.99600 (ima) 10.01360 (g) 10.00320 (e) -303.99300 (dataset) -303.01700 (and) -302.99400 (a) -303.98400 (ne) 15.01830 (w) -302.99500 (dataset) -303.01600 (of) -302.99600 (comple) 20.00890 (x) -303.99300 (ques\055) ] TJ T* [ (tions) -250.00600 (about) -250.01000 (abstr) 15.00730 (act) -249.99000 (shapes\056) ] TJ /R24 11.95520 Tf 36.91020 TL T* [ (1\056) -249.99000 (Intr) 18.01460 (oduction) ] TJ /R26 9.96260 Tf 11.95470 -19.27580 Td [ (This) -294.00700 (paper) -293.99300 (describes) -294.01600 (an) -293.98900 (approach) -293.98600 (to) -294.01700 (visual) -294.00200 (question) -293.99300 (an\055) ] TJ -11.95470 -11.95630 Td [ (swering) -340.98400 (based) -340.99500 (on) -342.01200 (a) -341.00700 (ne) 25.01670 (w) -341 (model) -341.01700 (architecture) -341 (that) -340.98700 (we) -341.98500 (call) -340.99700 (a) ] TJ /R32 9.96260 Tf 11.95510 TL T* [ (neur) 14.98950 (al) -298.01600 (module) -298.00800 (network) ] TJ /R26 9.96260 Tf 96.37070 0 Td [ (\050NMN\051\056) -298.01600 (This) -298.00700 (architecture) -297.98700 (mak) 10.01060 (es) -297.98700 (it) ] TJ -96.37070 -11.95470 Td [ (possible) -228.98300 (to) -229.98900 (answer) -229.01800 (natural) -229.98900 (language) -229.00600 (questions) -229.01800 (about) -229.98400 (images) ] TJ T* [ (using) -377.98700 (collections) -378.01800 (of) -378.01400 (jointly\055trained) -378.00200 (neural) -378.01000 (\223modules\224\054) -410.02000 (dy\055) ] TJ T* [ (namically) -230.99000 (composed) -230.01100 (into) -230.99600 (deep) -231.01900 (netw) 10.00940 (orks) -229.98900 (based) -231.01300 (on) -231.01000 (linguistic) ] TJ (structure\056) ' 11.95470 -12.30310 Td [ (Concretely) 64.99650 (\054) -376.99300 (gi) 24.98850 (v) 14.98280 (en) -351 (an) -351 (image) -350.98400 (and) -350.99400 (an) -352.01800 (associated) -350.98100 (question) ] TJ -11.95470 -11.95590 Td (\050e\056g\056) Tj /R32 9.96260 Tf 20.91090 0 Td [ (wher) 36.99440 (e) -321.99000 (is) -322.01100 (the) -322.01700 (do) 10.00810 (g\077) ] TJ /R26 9.96260 Tf 72.24800 0 Td [ (\051\054) -339.98400 (we) -321.98700 (wish) -322.00400 (to) -322.01200 (predict) -322.01200 (a) -321.99000 (correspond\055) ] TJ -93.15900 -11.95510 Td [ (ing) -288.01300 (answer) -287.98900 (\050e\056g\056) ] TJ /R32 9.96260 Tf 67.25200 0 Td [ (on) -287.98100 (the) -287.98400 (couc) 15.01830 (h) ] TJ /R26 9.96260 Tf 51.51290 0 Td [ (\054) -297.00700 (or) -287.99000 (perhaps) -287.98100 (just) ] TJ /R32 9.96260 Tf 67.73400 0 Td [ (couc) 15.01960 (h) ] TJ /R26 9.96260 Tf 23.64100 0 Td [ (\051) -287.99400 (\050) ] TJ ET Q 1 0 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 269.75600 176.64100 Tm (Fig\055) Tj -219.64400 -11.95470 Td [ (ure) -323.98800 (1) ] TJ ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 71.04340 164.68600 Tm [ (\051\056) -531.99000 (The) -323.98100 (visual) -323.99900 (question) -323.98900 (answering) -324.00500 (task) -324.01000 (has) -324.01700 <7369676e690263616e74> ] TJ -20.93120 -11.95510 Td [ <7369676e690263616e74> -285.98200 (applications) -285.98900 (to) -286.01800 (human\055robot) -285.99100 (interaction\054) -295.01900 (search\054) ] TJ 11.95510 TL T* [ (and) -239.99400 (accessibility) 65.01000 (\054) -243.01600 (and) -239.99300 (has) -239.99200 (been) -241.01700 (the) -239.99200 (subject) -240.01600 (of) -241.01600 (a) -240.00400 (great) -240 (deal) -241.01600 (of) ] TJ T* [ (recent) -337.00700 (research) -337.99500 (attention) -337.00600 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 155.90500 128.82100 Tm (3) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 160.88600 128.82100 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 166.73400 128.82100 Tm (10) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 176.69700 128.82100 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 182.54500 128.82100 Tm (26) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 192.50700 128.82100 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 198.36600 128.82100 Tm (28) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 208.32800 128.82100 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 214.17600 128.82100 Tm (33) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 224.13900 128.82100 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 229.98700 128.82100 Tm (40) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 239.94900 128.82100 Tm [ (\135\056) -571.98300 (The) -338.01700 (task) ] TJ -189.83700 -11.95590 Td [ (requires) -332.01500 (sophisticated) -332.99500 (understanding) -332.01800 (of) -333 (both) -332 (visual) -333.01600 (scenes) ] TJ 11.95510 TL T* [ (and) -287.98600 (natural) -286.99900 (language\056) -423.01000 (Recent) -288.01800 (successful) -286.99100 (approaches) -287.98600 (repre\055) ] TJ T* [ (sent) -263.98100 (questions) -263.99300 (as) -263.99400 (bags) -263.98300 (of) -263.99400 (w) 10 (ords\054) -268.01800 (or) -263.99400 (encode) -263.99300 (the) -263.98800 (question) -263.99800 (us\055) ] TJ T* [ (ing) -266.99700 (a) -267.99900 (recurrent) -266.98100 (neural) -267.99000 (netw) 10.00940 (ork) -267.00700 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 177.30400 81 Tm (28) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 187.26700 81 Tm [ (\135) -267.99800 (and) -267.00900 (train) -267.98200 (a) -267.01900 (simple) -268 (clas\055) ] TJ ET Q q 3088.62000 4427.13000 2362.50000 1355.19000 re W n /R62 Do Q q 10 0 0 10 0 0 cm BT /R26 8.96640 Tf 1 0 0 1 308.86200 415.11600 Tm [ (Figure) -245.01700 (1\072) -306.99700 (A) -245.02300 (sc) 0.98839 (hematic) -244.99800 (representation) -244.99600 (of) -243.98300 (our) -244.98500 (proposed) -244 (model\227the) ] TJ 10.95900 TL T* [ (shaded) -194.99100 (gray) -194.98800 (area) -194.98200 (is) -195.98700 (a) ] TJ /R32 8.96640 Tf 74.47190 0 Td [ (neur) 15.01370 (al) -194.99100 (module) -195.02100 (network) ] TJ /R26 8.96640 Tf 83.81210 0 Td [ (of) -195.01500 (the) -195.01000 (kind) -194.99300 (introduced) ] TJ -158.28400 -10.95900 Td [ (in) -294.01500 (this) -294.00100 (paper) 55.01500 (\056) -441.99300 (Our) -294.00400 (approach) -294.00700 (uses) -293.99000 (a) -293.99600 (natural) -294.01200 (language) -293.98500 (parser) -293.99600 (to) -294.01200 (dy\055) ] TJ T* [ (namically) -272 (lay) -271.99000 (out) -271.01200 (a) -271.99800 (deep) -272.01700 (netw) 9.99826 (ork) -271.99500 (composed) -272.00600 (of) -271.99500 (reusable) -270.99300 (modules\056) ] TJ 10.95820 TL T* [ (F) 15.00560 (or) -215.00900 (visual) -214.99300 (question) -214.99300 (answering) -215 (tasks\054) -221.98800 (an) -215.01700 (additional) -214.99500 (sequence) -214.97900 (model) ] TJ 10.95900 TL T* [ (pro) 15.01650 (vides) -249.98400 (sentence) -249.99700 (conte) 15 (xt) -250.01400 (and) -250 (learns) -250.01700 (common\055sense) -249.97800 (kno) 25.00660 (wledge\056) ] TJ /R26 9.96260 Tf 36.69690 TL T* [ <7369026572> -389.99400 (on) -390.98400 (the) -390.98600 (encoded) -390.00600 (question) -390.99800 (and) -390.00800 (image\056) -732.01500 (In) -390.01300 (contrast) -391.01300 (to) ] TJ 11.95510 TL T* [ (these) -310.01700 (monolithic) -310.01000 (approaches\054) -324.01500 (another) -310.01700 (line) -310.01200 (of) -309.98700 (w) 10.00320 (ork) -309.98300 (for) -309.98700 (te) 14.98280 (x\055) ] TJ 11.95590 TL T* [ (tual) -322.01000 (QA) -322.02000 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 347.92600 299.71400 Tm (23) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 357.88800 299.71400 Tm [ (\135) -321.98800 (and) -321 (image) -322.01000 (QA) -322.02000 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 430.46600 299.71400 Tm (27) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 440.42900 299.71400 Tm [ (\135) -321.99000 (uses) -322.01000 (semantic) -322 (parsers) -321.98500 (to) ] TJ -131.56600 -11.95510 Td [ (decompose) -302.98400 (questions) -301.98600 (into) -302.98400 (logical) -302.00400 (e) 15.01220 (xpressions\056) -468 (These) -303.00300 (logi\055) ] TJ 11.95510 TL T* [ (cal) -245.02000 (e) 15.01220 (xpressions) -244 (are) -244.99100 (e) 25.01050 (v) 24.98110 (aluated) -245.00800 (ag) 5.01877 (ainst) -244.99800 (a) -244.00300 (purely) -245.00300 (logical) -244.99300 (repre\055) ] TJ T* [ (sentation) -323.99000 (of) -323.98300 (the) -324.99700 (w) 10 (orld\054) -342.00400 (which) -324.99000 (may) -324.01000 (be) -323.98500 (pro) 14.98280 (vided) -324.00500 (directly) -324.98500 (or) ] TJ T* [ (e) 15.01220 (xtracted) -249.99000 (from) -249.98800 (an) -249.99700 (image) -249.98300 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 411.62600 251.89400 Tm (21) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 421.58800 251.89400 Tm (\135\056) Tj -100.77100 -13.71600 Td [ (In) -537.98600 (this) -536.98400 (paper) -537.99100 (we) -537.99100 (dra) 15.01220 (w) -537.98400 (from) -536.99900 (both) -538.00300 (lines) -538.00800 (of) -537.98900 (research\054) ] TJ -11.95510 -11.95510 Td [ (presenting) -353.99000 (a) -353.98500 (technique) -353.01700 (for) -353.98000 (inte) 14.99260 (grating) -354 (the) -354.01000 (representational) ] TJ 11.95590 TL T* [ (po) 24.98600 (wer) -371.00200 (of) -372.01400 (neural) -370.99200 (netw) 10.00810 (orks) -372.00200 (with) -371.01200 (the) -370.98700 <0365> 14.98770 (xible) -371.99700 (compositional) ] TJ 11.95510 TL T* [ (structure) -453.00800 (af) 25.00810 (forded) -451.99800 (by) -453.01300 (symbolic) -451.99800 (approaches) -453.01800 (to) -453.01300 (semantics\056) ] TJ T* [ (Rather) -395.99800 (than) -397.02000 (relying) -395.99000 (on) -396.00300 (a) -396.99700 (monolithic) -395.99300 (netw) 10.00810 (ork) -396.00700 (structure) -397.01700 (to) ] TJ T* [ (answer) -261.01300 (all) -261.00500 (questions\054) -263.01000 (our) -261.00800 (approach) -261.01000 (assembles) -260.98600 (a) -261.02000 (netw) 10.00810 (ork) -261.01000 (on) ] TJ T* [ (the) -196 <0379> -195.98700 (from) -195.99700 (a) -196.01100 (collection) -196.01400 (of) -196.00700 (specialized\054) -207.00500 (jointly\055learned) -196.00700 (mod\055) ] TJ T* [ (ules) -243.00400 (\050) ] TJ ET Q 1 0 0 rg q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 330.65000 154.49200 Tm [ (Figure) -242.99400 (1) ] TJ ET Q 0 g q 10 0 0 10 0 0 cm BT /R26 9.96260 Tf 1 0 0 1 364.06500 154.49200 Tm [ (\051\056) -307.98800 (Rather) -243.00400 (than) -243.98400 (using) -242.98900 (logic) -242.99900 (to) -243.00800 (reason) -243.01300 (o) 14.98280 (v) 14.98280 (er) -244.00800 (truth) ] TJ -55.20270 -11.95630 Td [ (v) 24.98110 (alues\054) -310.99500 (the) -299.00200 (representations) -299.01900 (computed) -298 (by) -299 (our) -299.00400 (model) -298.98400 (remain) ] TJ 11.95470 TL T* [ (entirely) -250.01700 (in) -249.98500 (the) -249.99000 (domain) -250.00700 (of) -249.99500 (visual) -250.01200 (features) -249.99700 (and) -249.99300 (attentions\056) ] TJ 11.95510 -13.71600 Td [ (Our) -245.01300 (approach) -245.99300 <02727374> -244.99300 (analyzes) -245.99300 (each) -244.98900 (question) -246.00300 (with) -244.99300 (a) -246.00300 (seman\055) ] TJ -11.95510 -11.95510 Td [ (tic) -211.01400 (parser) 39.98840 (\054) -218.98100 (and) -211.01900 (uses) -211.01100 (this) -211 (analysis) -210.99400 (to) -211.01400 (determine) -211.98400 (the) -211.01400 (basic) -211.01400 (com\055) ] TJ T* [ (putational) -210.98400 (units) -210.99400 (\050attention\054) -218.00800 <636c6173736902636174696f6e2c> -219.01200 (etc\056\051) -297.00900 (needed) -210.98400 (to) -211.01400 (an\055) ] TJ T* [ (swer) -204.98300 (the) -205.99700 (question\054) -214.00400 (as) -204.98300 (well) -205.01000 (as) -206.00500 (the) -205.01500 (relationships) -205.01500 (between) -206 (these) ] TJ ET Q Q Q q q 1 1 1 rg /a0 gs 48.40600 786.42200 515.18800 -52.69900 re f q /s5 gs /x6 Do Q q /s7 gs /x8 Do Q q /s9 gs /x10 Do Q q /s11 gs /x12 Do Q Q Q Q q 1 0 0 1 0 0 cm BT /F1 12 Tf 14.40000 TL ET 1 1 1 rg n 270 32 72 14 re f* 0.50000 0.50000 0.50000 rg BT /F2 9 Tf 10.80000 TL ET BT 1 0 0 1 301.50000 35 Tm (39) Tj T* ET Q endstream endobj 15 0 obj << /Filter /FlateDecode /Resources << /ExtGState << /a0 << /CA 1 /ca 1 >> >> /XObject << /x18 16 0 R >> >> /Length 28 /Group << /Type /Group /S /Transparency /CS /DeviceRGB /I true >> /BBox [ 78 746 96 765 ] /Type /XObject /Subtype /Form >> stream x+O4PH/VЯ0Pp 0 endstream endobj 16 0 obj << /Filter /FlateDecode /Resources 17 0 R /Length 107 /Type /XObject /BBox [ 78 746 96 765 ] /Subtype /Form >> stream xe AC̬wʠ =p,?]%+H-
Jc "82w8VSnGW;"
endstream
endobj
17 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
18 0 obj
<<
/Filter /FlateDecode
/BBox [ 4315.75000 4862.57000 4352.59000 4913.18000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 118
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x1@{bk
s>o(/P$B9$vNLOyL5PCz w)lB0[S6:quqWLKwY䲑䀗)ot^(ys%
endstream
endobj
19 0 obj
<<
/Type /ExtGState
/BM /Normal
>>
endobj
20 0 obj
<<
/S /Transparency
/Type /Group
/K true
/I true
>>
endobj
21 0 obj
<<
/Filter /FlateDecode
/BBox [ 3808.98000 4627.51000 3859.60000 4664.34000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 118
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x;@ާ¬vN
@"\E4'=J&5ba=|.d)fƥœUxAq\v^&Wƍ+tF&}
endstream
endobj
22 0 obj
<<
/Filter /FlateDecode
/BBox [ 3805.23000 5623.23000 3855.85000 5660.07000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 117
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x1@{bk
s>g(H
^ $B9$
hh%=IK&jj3zz *.YݘE7&n{沣4ǛDz:WH&s
endstream
endobj
23 0 obj
<<
/Filter /FlateDecode
/BBox [ 4724.07000 5136.82000 4760.91000 5187.43000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 120
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x1@Dާ¬ڜ I Pp} qhۏ_$-0
g>
>>
/Length 116
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x=AQFYW+Ɲ;VRx
+)x
w%br0?Aڪ`] #aIM\7
bބ-p%]q^QrVE2=/t=%e
endstream
endobj
25 0 obj
<<
/Filter /FlateDecode
/BBox [ 4724.10000 5136.78000 4760.94000 5187.39000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 119
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x̻
@|`}+ " L`0sH>d
5D0z`|d:
G*v"Xԁ#;J&W7:H8&[
endstream
endobj
26 0 obj
<<
/Filter /FlateDecode
/BBox [ 4724.30000 5552.18000 4761.14000 5602.80000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 118
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x
1"
$1 4h4%fmްq
>3Yw{
c¡&X~B
G9^TG땎t7ъ&
endstream
endobj
27 0 obj
<<
/Filter /FlateDecode
/BBox [ 4651.66000 5623.36000 4702.27000 5660.20000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 118
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x11}&NyG($$
jh'()f1j'ޑ:êk =IALMFhe60b4ig#x&
endstream
endobj
28 0 obj
<<
/Filter /FlateDecode
/BBox [ 4880.68000 5623.37000 4931.31000 5660.20000 ]
/Resources <<
/ExtGState <<
/R48 19 0 R
>>
>>
/Length 118
/Group 20 0 R
/Type /XObject
/FormType 1
/Matrix [ 1 0 0 1 0 0 ]
/Subtype /Form
>>
stream
x10ݧ?('2$i%fŒ^xv&j˅g>q2AJV"w֬)'?SZ|B7rN<|qS#>^&
endstream
endobj
29 0 obj
<<
/Filter /DCTDecode
/BitsPerComponent 8
/Height 333
/Length 17974
/ColorSpace /DeviceRGB
/Width 500
/Subtype /Image
>>
stream
Adobe d C
$, !$4.763.22:ASF:=N>22HbINVX]^]8EfmeZlS[]Y C**Y;2;YYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY M"
} !1AQa"q2#BR$3br
%&'()*456789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz
w !1AQaq"2B #3Rbr
$4%&'()*56789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz ? EJJk:iT*@jEZU"MhZVO JVRRlCeXFWIlL
i6ՂҴ\Erjrݴ\qHEK BE!)S,RRM"Jy4e%8iCҖi) m!bhM4iHh
41ҚC@
44
&ii( i4Jv)m!44,~Qx=5HW 4,l,t!5FmRG VGֱ&C{wnSTѪt~k.)#ԘKS-VFѪNQQ*ԪR- )vT)h6Th@)
UiT0[)Ld
ijJaZW+M"+L+@Xaj2(apiiL4M4CM4iCA4A4 ii i)3I
!4ii4ҸX
!iƚh IZLpIN`%4ғB?0Sh- 3T85k
R9\XX{RZ[ރi/{p*Xoy}!ț[Wb1aG_UݞKjGe]'չ/IʏMs̍j;Hb>bEc*A!N(b+nqS+:Z۱ӯ ^̈́ƺ%KCT
vFIVJWJVIVJIjtXIXI+)%ZXG1% 5H5 )-NSg3Q a@)x)
BQQڣj@B¢aSD1JS#L5#Tm@
4O4LCJi M0L4M!4@4SJi
4S(iJJSM&4bwugm_Ul'$N $Ҵb53Kk2]bik75Dw-v'ܾ[SJoyvc-*:U|2~UuUQv
b)CW+ILƋKcrjZ}_5sb&fjLSI7&))Ȥ"7)آb2*ŪK0ޢQo8ZOrsXC Zvq\.cn{*Fjֻ[|'cnc㱾-sj,` kM1CYNqfS$ՖZYHQՄVzfi=NZIdn$2XksnchMOVBԫqHFGY8[i12{L\DE1Fd3li=0ƣ&^2SBid)iKU in74viM&4f4 4 -6f
]{ap)i ,pUg,Jn;ɇ~u+^Z#>uљ# gU 8s~Onq1s'_Q*^ 'V69oɩ_YP XG?3H}:
M:/"(Hlh=)&=)1\Ml3b1J3bE0Iq!M4a4M"f(
=醟M56"aQZhoBc+1tAP 'NԽEj4pME Wq"l~ZEDn5P*XliT*U~hՙ?= *kv,@4)A3KX$2hk Fjw/ [Ψ')ana_7.Z1o 5(2 sVjrGthO|r)ްUu4&i>yej^=$w}ްMOAY2o{ =4.uG'+^Ἇ4մ?P