%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 ] /Type /Pages /Count 9 >> endobj 2 0 obj << /Title (Transparency by Design\072 Closing the Gap Between Performance and Interpretability in Visual Reasoning) /Producer (PyPDF2) /Author (David Mascharka\054 Philip Tran\054 Ryan Soklaski\054 Arjun Majumdar) /Subject (2018 IEEE Conference on Computer Vision and Pattern Recognition) >> endobj 3 0 obj << /Parent 1 0 R /Rotate 0 /Contents 13 0 R /Resources << /XObject << /x8 14 0 R /x10 17 0 R /x6 20 0 R /x12 23 0 R /R41 26 0 R /R40 27 0 R >> /ExtGState << /s9 28 0 R /s11 31 0 R /a0 << /CA 1 /ca 1 >> /R22 34 0 R /s5 35 0 R /s7 38 0 R >> /Font << /F2 41 0 R /R29 42 0 R /F1 45 0 R /R37 46 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R /R27 66 0 R /R31 70 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 74 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 75 0 R 76 0 R 77 0 R 78 0 R 79 0 R 80 0 R 81 0 R 82 0 R 83 0 R 84 0 R 85 0 R 86 0 R 87 0 R ] /Type /Page >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 88 0 R /Resources << /ExtGState << /R22 34 0 R >> /Font << /F2 89 0 R /F1 90 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 91 0 R 92 0 R 93 0 R 94 0 R 95 0 R 96 0 R 97 0 R 98 0 R 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 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 ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Contents 121 0 R /Resources << /XObject << /R113 122 0 R /R116 123 0 R /R114 124 0 R /R115 125 0 R >> /ExtGState << /R22 34 0 R >> /Font << /F2 126 0 R /F1 127 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 74 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 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 137 0 R 138 0 R 139 0 R 140 0 R 141 0 R 142 0 R 143 0 R 144 0 R 145 0 R 146 0 R 147 0 R 148 0 R 149 0 R 150 0 R 151 0 R 152 0 R 153 0 R 154 0 R ] /Type /Page >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 155 0 R /Resources << /ExtGState << /R22 34 0 R >> /Font << /F2 156 0 R /R130 157 0 R /R132 161 0 R /F1 165 0 R /R134 166 0 R /R136 169 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R /R128 173 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 176 0 R 177 0 R 178 0 R 179 0 R 180 0 R 181 0 R ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Contents 182 0 R /Resources << /ColorSpace << /R157 183 0 R >> /XObject << /R174 184 0 R /R175 185 0 R /R176 186 0 R /R155 205 0 R /R156 204 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] /Pattern << /R167 188 0 R /R166 195 0 R /R165 196 0 R /R164 197 0 R /R169 198 0 R /R168 199 0 R /R170 200 0 R /R171 201 0 R /R172 202 0 R /R173 203 0 R >> /ExtGState << /R154 187 0 R /R22 34 0 R >> /Font << /R181 206 0 R /R179 210 0 R /R177 213 0 R /F2 217 0 R /R132 161 0 R /F1 218 0 R /R136 169 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R >> /Shading << /R163 189 0 R >> >> /Group 74 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 219 0 R 220 0 R 221 0 R 222 0 R 223 0 R 224 0 R 225 0 R 226 0 R 227 0 R 228 0 R 229 0 R 230 0 R ] /Type /Page >> endobj 8 0 obj << /Parent 1 0 R /Rotate 0 /Contents 231 0 R /Resources << /XObject << /R202 232 0 R /R203 233 0 R /R204 234 0 R /R205 235 0 R >> /ExtGState << /R154 187 0 R /R22 34 0 R >> /Font << /F2 236 0 R /R206 237 0 R /R179 210 0 R /R177 213 0 R /R130 157 0 R /R132 161 0 R /F1 240 0 R /R136 169 0 R /R23 50 0 R /R25 59 0 R /R33 63 0 R /R31 70 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 74 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 241 0 R 242 0 R 243 0 R 244 0 R 245 0 R 246 0 R 247 0 R 248 0 R 249 0 R ] /Type /Page >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 250 0 R /Resources << /XObject << /R224 251 0 R /R225 252 0 R >> /ExtGState << /R22 34 0 R >> /Font << /R206 237 0 R /F2 253 0 R /F1 254 0 R /R23 50 0 R /R25 59 0 R /R33 63 0 R /R31 70 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 266 0 R /Resources << /XObject << /R242 267 0 R /R239 268 0 R >> /ExtGState << /R22 34 0 R >> /Font << /F2 269 0 R /R206 237 0 R /R130 157 0 R /R132 161 0 R /F1 270 0 R /R136 169 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /R33 63 0 R /R240 271 0 R /R128 173 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 275 0 R 276 0 R 277 0 R 278 0 R 279 0 R 280 0 R 281 0 R 282 0 R ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 283 0 R /Resources << /ExtGState << /R22 34 0 R >> /Font << /F2 284 0 R /R23 50 0 R /R35 55 0 R /R25 59 0 R /F1 285 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 12 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 13 0 obj << /Length 16767 >> stream q q q 0.1 0 0 0.1 0 0 cm /R22 gs 0 g q 10 0 0 10 0 0 cm BT /R23 14.3462 Tf 1 0 0 1 86.9691 675.067 Tm [ (T) 73.9912 (ranspar) 18.0074 (ency) -249.988 (by) -250.013 (Design\072) -309.994 (Closing) -249.99 (the) -250.008 (Gap) -250.012 (Between) -249.988 (P) 20.0061 (erf) 24.9889 (ormance) -249.998 (and) ] TJ 100.516 -17.9332 Td [ (Inter) 9.99625 (pr) 18.0048 (etability) -250.013 (in) -249.993 (V) 37.0137 (isual) -250.005 (Reasoning) ] TJ /R25 11.9552 Tf -88.7801 -37.8578 Td [ (Da) 19.9894 (vid) -250.003 (Mascharka) ] TJ /R27 7.9701 Tf 84.4148 4.33867 Td [ <03> -0.29866 ] TJ /R29 7.9701 Tf 4.98125 0 Td [ (1) -0.29866 ] TJ /R25 11.9552 Tf 31.632 -4.33867 Td [ (Philip) -250 (T) 34.9898 (ran) ] TJ /R29 7.9701 Tf 53.7137 4.33867 Td [ (2) -0.30019 ] TJ /R25 11.9552 Tf 31.6313 -4.33867 Td [ (Ryan) -249.995 (Soklaski) ] TJ /R29 7.9701 Tf 69.4109 4.33867 Td [ (1) -0.30019 ] TJ /R25 11.9552 Tf 31.6309 -4.33867 Td [ (Arjun) -250.014 (Majumdar) ] TJ /R27 7.9701 Tf 80.6852 4.33867 Td [ <03> -0.30019 ] TJ /R29 7.9701 Tf 4.98086 0 Td [ (1) -0.30019 ] TJ -336.457 -14.0578 Td [ (1) -0.29866 ] TJ /R25 11.9552 Tf 4.73203 -4.33789 Td [ (MIT) -250.005 (Lincoln) -249.985 (Laboratory) ] TJ /R27 7.9701 Tf 118.2 4.33789 Td [ (y) -0.19911 ] TJ /R29 7.9701 Tf 56.623 0 Td [ (2) -0.30019 ] TJ /R25 11.9552 Tf 4.73281 -4.33789 Td [ (Planck) -250 (Aerosystems) ] TJ /R27 7.9701 Tf 97.2902 4.33789 Td [ (z) -0.20217 ] TJ /R31 8.9664 Tf -256.879 -18.2859 Td [ (f) -0.8999 ] TJ /R33 8.9664 Tf 4.60781 0 Td (first\056last) Tj /R31 8.9664 Tf 53.798 0 Td [ (g) -0.90126 ] TJ /R33 8.9664 Tf 4.6082 0 Td [ (\100ll\056mit\056edu\054) -599.997 (phil\100planckaero\056com) ] TJ /R23 11.9552 Tf -97.0473 -41.0461 Td (Abstract) Tj /R35 9.9626 Tf -83.9277 -24.684 Td [ (V) 73.9913 (isual) -381.004 (question) -380.019 (answering) -380.987 (r) 37.0196 (equir) 36.9926 (es) -380.013 (high\055or) 37.0036 (der) -380.988 (r) 37.0183 (eason\055) ] TJ -11.9551 -11.9551 Td [ (ing) -487 (about) -486.01 (an) -487.007 (ima) 10.013 (g) 10.0032 (e) 9.99343 (\054) -545.985 (whic) 14.9987 (h) -487.012 (is) -485.984 (a) -487.012 (fundamental) -487.007 (capability) ] TJ 11.9551 TL T* [ (needed) -402.99 (by) -402.011 (mac) 14.9816 (hine) -403.019 (systems) -401.992 (to) -403.019 (follow) -402.005 (comple) 20.0089 (x) -402.996 (dir) 36.9914 (ectives\056) ] TJ 11.9559 TL T* [ (Recently) 55.0166 (\054) -346.998 (modular) -327.008 (networks) -327.989 (have) -327.003 (been) -327.982 (shown) -327.009 (to) -328.011 (be) -327.004 (an) -328.014 (ef\055) ] TJ 11.9551 TL T* [ (fective) -422.004 (fr) 14.9914 (ame) 14.9816 (work) -421.988 (for) -421.99 (performing) -422.002 (visual) -421.981 (r) 37.0183 (easoning) -422.013 (tasks\056) ] TJ T* [ (While) -288.014 (modular) -289.014 (networks) -287.996 (wer) 36.9889 (e) -287.996 (initially) -289 (designed) -288.006 (with) -287.999 (a) -289.006 (de\055) ] TJ T* [ (gr) 36.9852 (ee) -286.98 (of) -288.018 (model) -286.995 (tr) 14.9914 (anspar) 37.0061 (ency) 55.0019 (\054) -297.007 (their) -286.99 (performance) -286.982 (on) -287.981 (comple) 20.0089 (x) ] TJ T* [ (visual) -365.011 (r) 37.0196 (easoning) -363.983 (benc) 15.0183 (hmarks) -365.016 (was) -365.013 (lac) 20.004 (king) 14.9901 (\056) -654.006 (Curr) 37.0036 (ent) -364.991 (state\055) ] TJ T* [ (of\055the\055art) -393 (app) -1.00473 (r) 45.9985 (oac) 14.0038 (hes) -392.986 (do) -394.003 (not) -393.015 (pr) 44.9839 (o) 10.0032 (vide) -394.009 (an) -392.983 (ef) 18 (fective) -394.008 (mec) 15.0122 (ha\055) ] TJ 11.9559 TL T* [ (nism) -199.985 (for) -199.987 (under) 10.0155 (standing) -201.019 (the) -199.998 (r) 37.0196 (easoning) -200.011 (pr) 44.9851 (ocess\056) -293 (In) -200.984 (thi) 1 (s) -201.02 (paper) 111.018 (\054) ] TJ 11.9551 TL T* [ (we) -203.016 (close) -203.015 (the) -203.017 (performance) -202.996 (gap) -203.01 (between) -203.008 (interpr) 37.0061 (etable) -203.01 (models) ] TJ T* [ (and) -288.015 (state\055of\055the\055art) -287 (visual) -288.005 (r) 37.0196 (easoning) -287.996 (methods\056) -422.998 (W) 91.9871 (e) -287.996 (pr) 44.9851 (opose) ] TJ T* [ (a) -235.995 (set) -235.991 (of) -235.988 (visual\055r) 37.0036 (easoning) -236.985 (primit) 1.01576 (ives) -237.014 (whic) 15 (h\054) -239.012 (when) -235.98 (composed\054) ] TJ T* [ (manifest) -214.987 (as) -213.992 (a) -215.018 (model) -215.007 (capable) -215.005 (of) -213.991 (performing) -215.018 (comple) 20.0089 (x) -214.989 (r) 37.0183 (eason\055) ] TJ T* [ (ing) -280.015 (tasks) -278.992 (in) -280.019 (an) -279.983 (e) 19.9918 (xplicitly\055interpr) 36.9914 (etable) -280.016 (manner) 110.981 (\056) -398.99 (The) -280.017 <0264656c697479> ] TJ 11.9559 TL T* [ (and) -239.984 (interpr) 37.0067 (etability) -238.982 (of) -239.987 (the) -239.992 (primitives\047) -240.002 (outputs) -239.016 (enable) -239.984 (an) -239.989 (un\055) ] TJ 11.9551 TL T* [ (par) 14.9938 (alleled) -246.015 (ability) -246.003 (to) -245.006 (dia) 10.0155 (gnose) -245.981 (the) -245.991 (str) 36.9938 (engths) -246.013 (and) -245.983 (weaknesses) ] TJ T* [ (of) -263.983 (the) -265.007 (r) 37.0196 (esulting) -263.996 (model\056) -352.982 (Critically) 54.9835 (\054) -267.992 (we) -265.006 (show) -264.005 (that) -264.01 (these) -265.005 (prim\055) ] TJ T* [ (itives) -308.988 (ar) 36.9852 (e) -307.992 (highly) -308.983 (performant\054) -322.996 (ac) 15.0183 (hie) 14.9852 (ving) -308.995 (state\055of\055the\055art) -308.997 (ac\055) ] TJ T* [ (cur) 14.9846 (acy) -267.018 (of) -267.002 (99\0561\045) -266.995 (on) -267.985 (the) -267.007 (CLEVR) -267.014 (dataset\056) -361.003 (W) 91.9871 (e) -267.019 (also) -268.012 (show) -266.985 (that) ] TJ 11.9563 TL T* [ (our) -270.998 (model) -270.998 (is) -271 (able) -271.001 (to) -271.001 (ef) 18 (fectively) -271.006 (learn) -270.995 (g) 10.0032 (ener) 15.0196 (alized) -270.987 (r) 37.0183 (epr) 36.9816 (esen\055) ] TJ 11.9551 TL T* [ (tations) -360.015 (when) -360 (pr) 44.9839 (o) 10.0032 (vided) -358.99 (a) -360.013 (small) -359.99 (amount) -359.989 (of) -360.006 (data) -359.996 (containing) ] TJ T* [ (no) 10.0081 (vel) -341.983 (object) -342.005 (attrib) 19.9954 (utes\056) -586.984 (Us) 0.98881 (ing) -342.984 (the) -342.014 (CoGenT) -341.985 (g) 10.0032 (ener) 15.0196 (alization) ] TJ T* [ (task\054) -403.994 (we) -374.007 (show) -373.006 (mor) 36.9877 (e) -373.982 (than) -373.994 (a) -372.992 (20) -374.006 (per) 36.9816 (centa) 10.0057 (g) 10.0032 (e) -373.001 (point) -373.987 (impr) 44.9949 (o) 10.0032 (ve\055) ] TJ T* [ (ment) -249.987 (o) 10.0032 (ver) -250 (the) -249.99 (curr) 36.9865 (ent) -249.99 (state) -250.016 (of) -249.985 (the) -249.99 (art\056) ] TJ /R23 11.9552 Tf 38.1871 TL T* [ (1\056) -249.99 (Intr) 18.0146 (oduction) ] TJ /R25 9.9626 Tf 11.9551 -19.7027 Td [ (A) -307.985 (visual) -306.981 (question) -307.991 (answering) -308.007 (\050VQA\051) -307.985 (model) -306.983 (must) -307.998 (be) -307.988 (ca\055) ] TJ -11.9551 -11.9551 Td [ (pable) -291.002 (of) -291.989 (comple) 14.9987 (x) -291.005 (spatial) -292.005 (reasoning) -290.989 (o) 14.9828 (v) 14.9828 (er) -291.998 (an) -291.01 (image\056) -433.989 (F) 14.9926 (or) -291.988 (e) 15.0122 (x\055) ] TJ ET Q 3.98 w 0 G 501.121 1573.44 m 1446.11 1573.44 l S q 10 0 0 10 0 0 cm BT /R37 5.9776 Tf 1 0 0 1 60.141 150.742 Tm [ <03> -0.90058 ] TJ /R25 7.9701 Tf 4.3168 -2.81328 Td [ (Indicates) -249.984 (equal) -250.013 (contrib) 19.9966 (ution\056) ] TJ /R37 5.9776 Tf -3.92969 -6.98984 Td [ (y) -0.10006 ] TJ /R25 7.9701 Tf 3.92969 -2.81289 Td [ (This) -219.015 (material) -218.976 (is) -218.978 (based) -219 (upon) -218.987 (w) 10.0121 (ork) -218.011 (supported) -219.021 (by) -219.01 (the) -218.983 (Assistant) -219.01 (Secretary) ] TJ -14.3457 -9.46523 Td [ (of) -358.024 (Defense) -357.992 (for) -358.987 (Research) -357.995 (and) -357.999 (Engineering) -358.023 (under) -357.994 (Air) -359.017 (F) 15.0128 (orce) -358.009 (Contract) -358.003 (No\056) ] TJ 9.46406 TL T* [ (F) 74.0232 (A8721\05505\055C\0550002) -367.991 (and\057or) -367.986 (F) 74.0224 (A8702\05515\055D\0550001\056) -664.996 (An) 14.9975 (y) -367.989 (opinions\054) -397.984 <026e64696e67732c> ] TJ 9.46484 TL T* [ (conclusions) -215.987 (or) -215.989 (recommendations) -216.995 (e) 15.0036 (xpressed) -216.008 (in) -215.999 (this) -215.993 (material) -215.987 (are) -216.991 (those) -216.008 (of) -215.99 (the) ] TJ 9.46406 TL T* [ (author\050s\051) -212.98 (and) -211.994 (do) -212.981 (not) -213.022 (necessarily) -212.976 <726503656374> -212.012 (the) -213.002 (vie) 24.9973 (ws) -212.991 (of) -213 (the) -212.023 (Assistant) -212.982 (Secretary) ] TJ 9.46484 TL T* [ (of) -250.003 (Defense) -250.02 (for) -249.986 (Research) -250.024 (and) -249.978 (Engineering\056) ] TJ /R37 5.9776 Tf 10.416 -6.99023 Td [ (z) -0.10006 ] TJ /R25 7.9701 Tf 3.92969 -2.81289 Td [ (This) -249.99 (w) 10.0129 (ork) -250.016 (conducted) -249.997 (while) -249.984 (Philip) -249.994 (w) 10.0121 (as) -250.004 (at) -249.993 (MIT) -249.979 (Lincoln) -250 (Laboratory) 65.0166 (\056) ] TJ ET Q q 2362.37 0 0 1352.77 3088.62 4249.09 cm /R41 Do Q q 10 0 0 10 0 0 cm BT /R25 8.9664 Tf 1 0 0 1 308.862 413.95 Tm [ (Figure) -248.982 (1\056) -248.998 (A) -248.987 (diagram) -249.012 (of) -249.992 (a) -248.995 (visual) -249.014 (question) -249.02 (answering) -248.982 (task\054) -249.004 (in) -249.009 (which) ] TJ 10.959 TL T* [ (our) -393.017 (proposed) -391.988 (T) 35.0212 (ransparenc) 15.0165 (y) -393.004 (by) -393.004 (Design) -392.018 (netw) 9.99826 (ork) -393.02 (\050TbD\055net\051) -391.98 (com\055) ] TJ T* [ (poses) -342 (a) -343.01 (series) -341.978 (of) -342.002 (attention) -342.993 (masks) -341.989 (that) -341.989 (allo) 25.0066 (w) -342 (it) -343.018 (to) -341.978 (correctly) -342.01 (count) ] TJ T* [ (tw) 10.02 (o) -249.978 (lar) 18.0116 (ge) -249.997 (metal) -249.989 (c) 15.011 (ylinders) -249.989 (in) -250.014 (the) -249.989 (image\056) ] TJ /R25 9.9626 Tf 35.316 TL T* [ (ample\054) -354.002 (in) -334.01 (order) -333.983 (to) -332.991 (answer) -333.983 (the) -332.993 (question) -333.988 (\223What) -334.018 (color) -332.993 (is) -334.008 (the) ] TJ 11.9551 TL T* [ (cube) -275.991 (to) -275.981 (the) -275.986 (right) -276.01 (of) -275.991 (the) -275.986 (lar) 17.997 (ge) -275.993 (metal) -274.993 (sphere\077\224\054) -283.012 (a) -275.998 (model) -276.008 (must) ] TJ T* [ (identify) -417 (which) -415.994 (sphere) -416.984 (is) -417.016 (the) -416 (lar) 17.997 (ge) -416.989 (metal) -415.989 (one\054) -459.012 (understand) ] TJ T* [ (what) -231.01 (it) -231.006 (means) -231.006 (for) -230.984 (an) -231.02 (object) -231.006 (to) -231.01 (be) -232 (to) -231.01 (the) -231.01 (right) -231 (of) -231.02 (another) 40.0031 (\054) -235.015 (and) ] TJ 11.9563 TL T* [ (apply) -300.011 (this) -300.004 (concept) -300.999 (s) 0.98513 (patially) -301.018 (to) -300.019 (the) -299.979 (attended) -300.014 (sphere\056) -461.016 (W) 39.9933 (ithin) ] TJ 11.9547 TL T* [ (this) -207 (ne) 25.0154 (w) -206.983 (re) 15.0098 (gion) -205.983 (of) -206.983 (interest\054) -216.018 (the) -207.014 (model) -206.02 (must) -206.995 <026e64> -207 (the) -207.014 (cube) -206.985 (and) ] TJ T* [ (determine) -268.994 (its) -270.014 (color) 54.981 (\056) -368.015 (This) -268.992 (beha) 19.9967 (vior) -270.019 (should) -268.98 (be) -269.994 (compositional) ] TJ T* [ (to) -249.985 (allo) 24.9909 (w) -249.995 (for) -250 (arbitrarily) -250.017 (long) -250.012 (reasoning) -250.017 (chains\056) ] TJ 11.9551 -13.6988 Td [ (While) -286.996 (a) -285.996 (wide) -287.001 (v) 24.9811 (ariety) -287.004 (of) -285.991 (models) -287.021 (ha) 19.9967 (v) 14.9828 (e) -285.996 (recently) -287.001 (been) -287.011 (pro\055) ] TJ -11.9551 -11.9551 Td [ (posed) -245 (for) -245.02 (the) -245.011 (VQA) -245.011 (task) -243.984 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 409.036 236.417 Tm (6) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 414.018 236.417 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 418.949 236.417 Tm (10) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 428.912 236.417 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 433.843 236.417 Tm (19) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 443.806 236.417 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 448.738 236.417 Tm (21) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 458.7 236.417 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 463.632 236.417 Tm (28) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 473.594 236.417 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 478.526 236.417 Tm (30) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 488.488 236.417 Tm [ (\135\054) -245.998 (neural) -245.013 (mod\055) ] TJ -179.626 -11.9551 Td [ (ule) -376.989 (netw) 10.0081 (orks) -376.003 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 368.279 224.462 Tm (2) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 373.26 224.462 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 379.507 224.462 Tm (3) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 384.488 224.462 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 390.725 224.462 Tm (10) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 400.687 224.462 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 406.934 224.462 Tm (14) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 416.896 224.462 Tm [ (\135) -376.021 (are) -377.006 (among) -377.011 (the) -376.006 (most) -377.006 (intuiti) 24.9811 (v) 14.9828 (e\056) ] TJ -108.034 -11.9551 Td [ (Introduced) -338.005 (by) -338.012 (Andreas) ] TJ /R35 9.9626 Tf 96.4238 0 Td [ (et) -338.017 (al) ] TJ /R25 9.9626 Tf 18.323 0 Td [ (\056) -338.017 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 432.785 212.507 Tm (2) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 437.766 212.507 Tm [ (\135\054) -360.999 (neural) -338.017 (module) -337.993 (netw) 10.0081 (orks) ] TJ -128.904 -11.9551 Td [ (compose) -284.987 (a) -283.997 <7175657374696f6e2d73706563690263> -284.997 (ne) 0.98023 (u) -1.01454 (r) 1.01454 (al) -285.006 (netw) 10.0081 (ork\054) -293 (dra) 15.0122 (wing) -284.992 (from) ] TJ 11.9559 TL T* [ (a) -428.012 (set) -429.017 (of) -428.005 (modules) -428.009 (that) -427.992 (each) -428.997 (perform) -427.997 (an) -428.007 (indi) 25 (vidual) -429.002 (opera\055) ] TJ 11.9551 TL T* [ (tion\056) -510.006 (This) -316.004 (design) -316.981 (closely) -316.006 (models) -317.016 (the) -316.016 (compositional) -316.981 (nature) ] TJ T* [ (of) -390.991 (visual) -391.006 (reasoning) -391.011 (tasks\056) -732.01 (In) -390.994 (the) -390.984 (original) -391.003 (w) 10.0032 (ork\054) -425.998 (modules) ] TJ T* [ (were) -228.011 (designed) -228.996 (with) -228.018 (an) -228.981 (attention) -228.006 (mechanism\054) -232.981 (which) -229.006 (allo) 24.9909 (wed) ] TJ T* [ (for) -394.995 (insight) -394.01 (into) -395.008 (the) -394.986 (model\047) 54.9908 (s) -394.988 (operation\056) -744.008 (Ho) 24.986 (we) 25.0154 (v) 14.9828 (er) 39.9835 (\054) -430.982 (the) -394.983 (ap\055) ] TJ T* [ (proach) -348.018 (did) -348.983 (not) -348.003 (perform) -348.993 (well) -348.006 (on) -348.991 (comple) 14.9975 (x) -348.015 (visual) -349.015 (reasoning) ] TJ 11.9559 TL T* [ (tasks) -373.996 (such) -374.004 (as) -372.994 (CLEVR) -374.018 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 406.217 116.865 Tm (13) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 416.179 116.865 Tm [ (\135\056) -680.984 <4d6f646902636174696f6e73> -373.992 (by) -374.006 (Johnson) ] TJ /R35 9.9626 Tf 121.74 0 Td (et) Tj -229.057 -11.9551 Td (al) Tj /R25 9.9626 Tf 7.76094 0 Td [ (\056) -291.008 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 325.33 104.91 Tm (14) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R25 9.9626 Tf 1 0 0 1 335.293 104.91 Tm [ (\135) -291.013 (address) -291 (the) -291.983 (performance) -291.01 (issue) -290.991 (at) -291.005 (the) -291 (cost) -292.015 (of) -291.01 (losing) ] TJ -26.4309 -11.9551 Td [ (model) -296.985 (transparenc) 14.9828 (y) 65.0137 (\056) -450 (This) -296.985 (is) -296.995 (problematic\054) -307.993 (because) -297.009 (the) -297 (abil\055) ] TJ 11.9551 TL T* [ (ity) -326.985 (to) -328.011 (inspect) -326.982 (each) -327.012 (step) -328.009 (of) -327.004 (the) -328.014 (reasoning) -326.985 (process) -326.994 (is) -328.009 (crucial) ] TJ ET Q Q Q q q 1 1 1 rg /a0 gs 48.406 786.422 515.188 -52.699 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 14.4 TL ET 1 1 1 rg n 270 32 72 14 re f* 0.5 0.5 0.5 rg BT /F2 9 Tf 10.8 TL ET BT 1 0 0 1 297 35 Tm (4942) Tj T* ET Q endstream endobj 14 0 obj << /Filter /FlateDecode /Resources << /ExtGState << /a0 << /CA 1 /ca 1 >> >> /XObject << /x18 15 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 15 0 obj << /Filter /FlateDecode /Resources 16 0 R /Length 107 /Type /XObject /BBox [ 78 746 96 765 ] /Subtype /Form >> stream xe AC̬wʠ =p,?]%+H-
Jc "82w8VSnGW;"
endstream
endobj
16 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
17 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x21 18 0 R
>>
>>
/Length 28
/Group <<
/Type /Group
/S /Transparency
/CS /DeviceRGB
/I true
>>
/BBox [ 96 752 110 775 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ02Tp
0w
endstream
endobj
18 0 obj
<<
/Filter /FlateDecode
/Resources 19 0 R
/Length 118
/Type /XObject
/BBox [ 96 752 110 775 ]
/Subtype /Form
>>
stream
xe=QSp#lZq|ddLB:D[c¸>*FI&(\T=s&GuJѩIY\I}#
.!.
endstream
endobj
19 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
20 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x15 21 0 R
>>
>>
/Length 28
/Group <<
/Type /Group
/S /Transparency
/CS /DeviceRGB
/I true
>>
/BBox [ 67 752 84 775 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ04Up
0
endstream
endobj
21 0 obj
<<
/Filter /FlateDecode
/Resources 22 0 R
/Length 228
/Type /XObject
/BBox [ 67 752 84 775 ]
/Subtype /Form
>>
stream
xeQKn!s ?FPav6R٪TS.
b];15YyR
{7QL.\:Rv/x9l+L7h%1!}i/AI(kz"U&,YO![R hg{3}4/GyYF:!w}Gn+'xJcO9i뽼_-:`
endstream
endobj
22 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
23 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x24 24 0 R
>>
>>
/Length 28
/Group <<
/Type /Group
/S /Transparency
/CS /DeviceRGB
/I true
>>
/BBox [ 132 751 480 772 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ02Qp
0
endstream
endobj
24 0 obj
<<
/Filter /FlateDecode
/Resources 25 0 R
/Length 53223
/Type /XObject
/BBox [ 132 751 480 772 ]
/Subtype /Form
>>
stream
xtI:6%Q㨈?7rA= u%6 ?Y(WbWo{B>9
x`Znϳ|8{3?0x*z ǃ|,@:w>`c|*ϻⳅKO3`g
:_|}}><.6`Z{{3]#<_o"~:ͺgk7/Ұ@|K yp ]03ʷCmş8˽Y?>(3!Bwqs.Z8,~~=rMT̩y+/*w: uBZ_`ߵp`%M?ɝ1ɳw=vDۉy&xb4Q>d@ sg~lA