| 1 | <?xml version='1.0' encoding='UTF-8'?> |
| 2 | <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1280 720" width="1280" height="720"> |
| 3 | <defs> |
| 4 | <filter id="cardShadow" x="-10%" y="-10%" width="120%" height="120%"> |
| 5 | <feGaussianBlur in="SourceAlpha" stdDeviation="4" /> |
| 6 | <feOffset dx="0" dy="3" /> |
| 7 | <feFlood flood-color="#1A365D" flood-opacity="0.07" /> |
| 8 | <feComposite in2="SourceAlpha" operator="in" /> |
| 9 | <feMerge><feMergeNode /><feMergeNode in="SourceGraphic" /></feMerge> |
| 10 | </filter> |
| 11 | </defs> |
| 12 | |
| 13 | <g id="bg"> |
| 14 | <rect width="1280" height="720" fill="#FFFFFF" /> |
| 15 | </g> |
| 16 | |
| 17 | <g id="header"> |
| 18 | <rect x="60" y="56" width="6" height="40" fill="#3182CE" /> |
| 19 | <text x="84" y="88" font-family="Georgia, "Microsoft YaHei", serif" font-size="36" font-weight="700" fill="#1A365D">A simpler model, better results, less training</text> |
| 20 | <text x="84" y="118" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">WHY THIS PAPER MATTERS · 2017 SOTA ON WMT'14 TRANSLATION</text> |
| 21 | </g> |
| 22 | |
| 23 | <g id="kpi-bleu-ende"> |
| 24 | <path fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)" d="M72,160 H333 A12,12 0 0 1 345,172 V388 A12,12 0 0 1 333,400 H72 A12,12 0 0 1 60,388 V172 A12,12 0 0 1 72,160 Z" /> |
| 25 | <path fill="#1A365D" d="M63,160 H63 A3,3 0 0 1 66,163 V397 A3,3 0 0 1 63,400 H63 A3,3 0 0 1 60,397 V163 A3,3 0 0 1 63,160 Z" /> |
| 26 | |
| 27 | <g transform="translate(80.0, 184.0) scale(1.3333333333333333)" fill="none" stroke="#1A365D" stroke-width="2"> |
| 28 | <path d="M8 21l8 0" /> |
| 29 | <path d="M12 17l0 4" /> |
| 30 | <path d="M7 4l10 0" /> |
| 31 | <path d="M17 4v8a5 5 0 0 1 -10 0v-8" /> |
| 32 | <path d="M3 9a2 2 0 1 0 4 0a2 2 0 1 0 -4 0" /> |
| 33 | <path d="M17 9a2 2 0 1 0 4 0a2 2 0 1 0 -4 0" /> |
| 34 | </g> |
| 35 | <text x="80" y="246" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1">EN → DE</text> |
| 36 | <text x="80" y="270" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" fill="#4A5568">BLEU on WMT'14</text> |
| 37 | <text x="80" y="336" font-family="Georgia, "Microsoft YaHei", serif" font-size="64" font-weight="700" fill="#1A365D">28.4</text> |
| 38 | |
| 39 | <g transform="translate(80.0, 354.0) scale(0.8333333333333334)" fill="none" stroke="#2F855A" stroke-width="2"> |
| 40 | <path d="M17 7l-10 10" /> |
| 41 | <path d="M8 7l9 0l0 9" /> |
| 42 | </g> |
| 43 | <text x="106" y="370" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="16" font-weight="700" fill="#2F855A">+2.0</text> |
| 44 | <text x="150" y="370" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" fill="#4A5568">over prior ensembles</text> |
| 45 | </g> |
| 46 | |
| 47 | <g id="kpi-bleu-enfr"> |
| 48 | <path fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)" d="M377,160 H638 A12,12 0 0 1 650,172 V388 A12,12 0 0 1 638,400 H377 A12,12 0 0 1 365,388 V172 A12,12 0 0 1 377,160 Z" /> |
| 49 | <path fill="#3182CE" d="M368,160 H368 A3,3 0 0 1 371,163 V397 A3,3 0 0 1 368,400 H368 A3,3 0 0 1 365,397 V163 A3,3 0 0 1 368,160 Z" /> |
| 50 | |
| 51 | <g transform="translate(385.0, 184.0) scale(1.3333333333333333)" fill="none" stroke="#3182CE" stroke-width="2"> |
| 52 | <path d="M8 21l8 0" /> |
| 53 | <path d="M12 17l0 4" /> |
| 54 | <path d="M7 4l10 0" /> |
| 55 | <path d="M17 4v8a5 5 0 0 1 -10 0v-8" /> |
| 56 | <path d="M3 9a2 2 0 1 0 4 0a2 2 0 1 0 -4 0" /> |
| 57 | <path d="M17 9a2 2 0 1 0 4 0a2 2 0 1 0 -4 0" /> |
| 58 | </g> |
| 59 | <text x="385" y="246" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1">EN → FR</text> |
| 60 | <text x="385" y="270" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" fill="#4A5568">BLEU on WMT'14</text> |
| 61 | <text x="385" y="336" font-family="Georgia, "Microsoft YaHei", serif" font-size="64" font-weight="700" fill="#3182CE">41.8</text> |
| 62 | |
| 63 | <g transform="translate(385.0, 354.0) scale(0.8333333333333334)" fill="none" stroke="#2F855A" stroke-width="2"> |
| 64 | <path d="M12 17.75l-6.172 3.245l1.179 -6.873l-5 -4.867l6.9 -1l3.086 -6.253l3.086 6.253l6.9 1l-5 4.867l1.179 6.873l-6.158 -3.245" /> |
| 65 | </g> |
| 66 | <text x="411" y="370" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#2F855A">new single-model SOTA</text> |
| 67 | </g> |
| 68 | |
| 69 | <g id="kpi-training"> |
| 70 | <path fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)" d="M682,160 H943 A12,12 0 0 1 955,172 V388 A12,12 0 0 1 943,400 H682 A12,12 0 0 1 670,388 V172 A12,12 0 0 1 682,160 Z" /> |
| 71 | <path fill="#1A365D" d="M673,160 H673 A3,3 0 0 1 676,163 V397 A3,3 0 0 1 673,400 H673 A3,3 0 0 1 670,397 V163 A3,3 0 0 1 673,160 Z" /> |
| 72 | |
| 73 | <g transform="translate(690.0, 184.0) scale(1.3333333333333333)" fill="none" stroke="#1A365D" stroke-width="2"> |
| 74 | <path d="M13 3l0 7l6 0l-8 11l0 -7l-6 0l8 -11" /> |
| 75 | </g> |
| 76 | <text x="690" y="246" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1">TRAINING TIME</text> |
| 77 | <text x="690" y="270" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" fill="#4A5568">Transformer (big)</text> |
| 78 | <text x="690" y="336" font-family="Georgia, "Microsoft YaHei", serif" font-size="48" font-weight="700" fill="#1A365D">3.5 days</text> |
| 79 | |
| 80 | <g transform="translate(690.0, 354.0) scale(0.8333333333333334)" fill="none" stroke="#4A5568" stroke-width="2"> |
| 81 | <path d="M5 6a1 1 0 0 1 1 -1h12a1 1 0 0 1 1 1v12a1 1 0 0 1 -1 1h-12a1 1 0 0 1 -1 -1l0 -12" /> |
| 82 | <path d="M9 9h6v6h-6l0 -6" /> |
| 83 | <path d="M3 10h2" /> |
| 84 | <path d="M3 14h2" /> |
| 85 | <path d="M10 3v2" /> |
| 86 | <path d="M14 3v2" /> |
| 87 | <path d="M21 10h-2" /> |
| 88 | <path d="M21 14h-2" /> |
| 89 | <path d="M14 21v-2" /> |
| 90 | <path d="M10 21v-2" /> |
| 91 | </g> |
| 92 | <text x="716" y="370" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" fill="#4A5568">on 8× P100 GPUs</text> |
| 93 | </g> |
| 94 | |
| 95 | <g id="kpi-no-rnn-cnn"> |
| 96 | <path fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)" d="M987,160 H1208 A12,12 0 0 1 1220,172 V388 A12,12 0 0 1 1208,400 H987 A12,12 0 0 1 975,388 V172 A12,12 0 0 1 987,160 Z" /> |
| 97 | <path fill="#3182CE" d="M978,160 H978 A3,3 0 0 1 981,163 V397 A3,3 0 0 1 978,400 H978 A3,3 0 0 1 975,397 V163 A3,3 0 0 1 978,160 Z" /> |
| 98 | |
| 99 | <g transform="translate(995.0, 184.0) scale(1.3333333333333333)" fill="none" stroke="#3182CE" stroke-width="2"> |
| 100 | <path d="M10 12a2 2 0 1 0 4 0a2 2 0 0 0 -4 0" /> |
| 101 | <path d="M21 12c-2.4 4 -5.4 6 -9 6c-3.6 0 -6.6 -2 -9 -6c2.4 -4 5.4 -6 9 -6c3.6 0 6.6 2 9 6" /> |
| 102 | </g> |
| 103 | <text x="995" y="246" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1">ARCHITECTURE</text> |
| 104 | <text x="995" y="270" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" fill="#4A5568">building blocks</text> |
| 105 | <text x="995" y="316" font-family="Georgia, "Microsoft YaHei", serif" font-size="40" font-weight="700" fill="#1A365D">Attention</text> |
| 106 | <text x="995" y="346" font-family="Georgia, "Microsoft YaHei", serif" font-size="22" font-style="italic" fill="#4A5568">only</text> |
| 107 | <text x="995" y="376" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" fill="#4A5568">no recurrence · no convolution</text> |
| 108 | </g> |
| 109 | |
| 110 | <g id="takeaway-band"> |
| 111 | <path fill="#F5F7FA" stroke="#E2E8F0" stroke-width="1" d="M72,450 H1208 A12,12 0 0 1 1220,462 V598 A12,12 0 0 1 1208,610 H72 A12,12 0 0 1 60,598 V462 A12,12 0 0 1 72,450 Z" /> |
| 112 | <path fill="#1A365D" d="M63,450 H63 A3,3 0 0 1 66,453 V607 A3,3 0 0 1 63,610 H63 A3,3 0 0 1 60,607 V453 A3,3 0 0 1 63,450 Z" /> |
| 113 | <text x="100" y="490" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#3182CE" letter-spacing="1.5">THE HEADLINE</text> |
| 114 | <text x="100" y="520" font-family="Georgia, "Microsoft YaHei", serif" font-size="22" font-weight="700" fill="#1A365D"> |
| 115 | The Transformer reaches <tspan fill="#3182CE">new translation SOTA</tspan> while training |
| 116 | </text> |
| 117 | <text x="100" y="548" font-family="Georgia, "Microsoft YaHei", serif" font-size="22" font-weight="700" fill="#1A365D"> |
| 118 | in <tspan fill="#3182CE">a fraction of the time</tspan> of strong baselines. |
| 119 | </text> |
| 120 | <text x="100" y="580" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="16" fill="#4A5568"> |
| 121 | Quality up · cost down · parallelization unlocked. The whole rest of the paper explains |
| 122 | </text> |
| 123 | <text x="100" y="602" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="16" fill="#4A5568"> |
| 124 | how attention alone makes this possible — and why nothing else is needed. |
| 125 | </text> |
| 126 | </g> |
| 127 | |
| 128 | <g id="footer"> |
| 129 | <text x="60" y="694" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">Source: Vaswani et al. 2017, Table 2 (BLEU) and §5.2 (training cost)</text> |
| 130 | <text x="1220" y="694" text-anchor="end" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">02</text> |
| 131 | </g> |
| 132 | </svg> |