File size: 89,544 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:13:19.871174Z"
},
"title": "Finetuning Pretrained Transformers into Variational Autoencoders",
"authors": [
{
"first": "Seongmin",
"middle": [],
"last": "Park",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "ActionPower Seoul",
"location": {
"country": "Republic of Korea"
}
},
"email": "seongmin.park@actionpower.kr"
},
{
"first": "Jihwa",
"middle": [],
"last": "Lee",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "ActionPower Seoul",
"location": {
"country": "Republic of Korea"
}
},
"email": "jihwa.lee@actionpower.kr"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Text variational autoencoders (VAEs) are notorious for posterior collapse, a phenomenon where the model's decoder learns to ignore signals from the encoder. Because posterior collapse is known to be exacerbated by expressive decoders, Transformers have seen limited adoption as components of text VAEs. Existing studies that incorporate Transformers into text VAEs (Li et al., 2020; Fang et al., 2021) mitigate posterior collapse using massive pretraining, a technique unavailable to most of the research community without extensive computing resources. We present a simple two-phase training scheme to convert a sequence-to-sequence Transformer into a VAE with just finetuning. The resulting language model is competitive with massively pretrained Transformer-based VAEs in some internal metrics while falling short on others. To facilitate training we comprehensively explore the impact of common posterior collapse alleviation techniques in the literature. We release our code for reproducability 1 .",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Text variational autoencoders (VAEs) are notorious for posterior collapse, a phenomenon where the model's decoder learns to ignore signals from the encoder. Because posterior collapse is known to be exacerbated by expressive decoders, Transformers have seen limited adoption as components of text VAEs. Existing studies that incorporate Transformers into text VAEs (Li et al., 2020; Fang et al., 2021) mitigate posterior collapse using massive pretraining, a technique unavailable to most of the research community without extensive computing resources. We present a simple two-phase training scheme to convert a sequence-to-sequence Transformer into a VAE with just finetuning. The resulting language model is competitive with massively pretrained Transformer-based VAEs in some internal metrics while falling short on others. To facilitate training we comprehensively explore the impact of common posterior collapse alleviation techniques in the literature. We release our code for reproducability 1 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Properly tamed latent models offer explainable and interpolatable representations of observed data. Recent works have shown such models to be especially useful in unsupervised learning settings. adapt a generative latent text model for successful unsupervised text style transfer and machine translation. achieve superior language modeling performance against common conditional counterparts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A popular variant of deep latent models is the variational autoencoder (VAE) (Kingma and Welling, 2014) . For each observed x, the model assumes the existence of a corresponding multidimensional latent vector z. Since the log evidence log p(x) is intractable for most interesting problems, the training process for VAEs opts instead to minimize the log evidence lower bound (ELBO): [log(p(x|z) )]\u2212D KL (q(z|x)||p(z)) (1) q(z|x) is a tractable, assumed posterior commonly modeled with a parametrized encoder q \u03c6 (z|x), while p(x|z) is the likelihood parametrized with a decoder p \u03b8 (x|z) that optimizes against reconstruction loss. While effective in theory, a common empirical challenge VAEs present during training is posterior collapse -a phenomenon where the decoder ignores the latent signal from z (and thus the originating input) during reconstruction. Posterior collapse can be diagnosed by checking if D KL (q(z|x)||p(z)) tends to zero during training.",
"cite_spans": [
{
"start": 77,
"end": 103,
"text": "(Kingma and Welling, 2014)",
"ref_id": "BIBREF9"
},
{
"start": 382,
"end": 393,
"text": "[log(p(x|z)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "E z\u223cq(z|x)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "After Bowman et al. (2016) adopted VAE for text, subsequent studies have been introduced with attempts to mitigate posterior collapse in VAE language models (LMs). However, the brittle training process of VAE LMs remains an unsolved problem. present a method to utilize deep Transformer (Vaswani et al., 2017) models as components of VAE LMs. Transformer-based VAEs tap into the state-of-the-art capabilities of Transformers while retaining representational advantages of VAE LMs. The paper mitigates posterior collapse by massive pretraining and a cyclical annealing schedule (Fu et al., 2019) .",
"cite_spans": [
{
"start": 6,
"end": 26,
"text": "Bowman et al. (2016)",
"ref_id": "BIBREF0"
},
{
"start": 287,
"end": 309,
"text": "(Vaswani et al., 2017)",
"ref_id": "BIBREF21"
},
{
"start": 577,
"end": 594,
"text": "(Fu et al., 2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "While the study presents a promising outlook for Transformer VAEs, the suggested method is not accessible to researchers who lack access to large, target-domain-specific corpora or the computing power for massive LM pretraining. Therefore, a demand arises for a way to finetune an existing Transformer model into a VAE LM with limited resources. Our research attempts to fill this gap in the literature, and makes the following contributions:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 We present a simple but reliable (as replicated across several datasets) scheme to teach latent structure to a pretrained Transformer model by just finetuning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 We convert a pretrained sequence-to-sequence Transformer into a VAE, instead of using two separate encoder-only (Devlin et al., 2019) and decoder-only (Radford et al., 2019) Transformers as in previous literature. This eliminates the need to maintain separate tokenizers and configurations for encoder and decoder.",
"cite_spans": [
{
"start": 114,
"end": 135,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF2"
},
{
"start": 153,
"end": 175,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 We conduct ablation studies and extensive experiments to gauge the effectiveness of commonly used posterior collapse mitigation methods in taming Transformer VAEs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The resulting model extends existing Transformer architectures and can be initialized from pretrained non-latent model checkpoints.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Most VAE LMs employ recurrent neural networks (RNNs) as encoders and decoders. This is in part because enforcing a latent bottleneck layer undermines the effectiveness of encoder-decoder cross-attention in Transformers, and in significant part due to the co-occurrence of posterior collapse and powerful and deep decoder layers. overcome such training difficulties by massively increasing the number of training samples (104,213,036 sentences) for LM pretraining. and Fang et al. (2021) also finds success with Transformer VAEs for text generation. To avoid posterior collapse, Fang et al. (2021) follow the exact cyclic KL mitigation approach as that of , while introduce noise to network input.",
"cite_spans": [
{
"start": 468,
"end": 486,
"text": "Fang et al. (2021)",
"ref_id": "BIBREF3"
},
{
"start": 578,
"end": 596,
"text": "Fang et al. (2021)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background 2.1 Transformer Text VAEs",
"sec_num": "2"
},
{
"text": "This study identifies and explores the effect of popular posterior collapse mitigation methods in low-resource Transformer VAE training. We do not examine importance-weighted autoencoders (Burda et al., 2016) and semi-amortized autoencoders (Kim et al., 2018) to limit the scope of our experiments to unsophisticated prior distributions.",
"cite_spans": [
{
"start": 188,
"end": 208,
"text": "(Burda et al., 2016)",
"ref_id": "BIBREF1"
},
{
"start": 241,
"end": 259,
"text": "(Kim et al., 2018)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Techniques to mitigate posterior collapse",
"sec_num": "2.2"
},
{
"text": "Bowman et al. (2016) increases the KL term of the ELBO from zero to its full value during early stages of training, where the decoder learns to simply treat latent signal z as noise. Fu et al. (2019) extend this technique by cyclically manipulating the weight of the KL term. \u03b2-VAE (Higgins et al., 2017) and Yan et al. (2020) adopt a similar approach. We train the network without the KL term of the ELBO and retain encoder weights before jointly training the whole network. (Shen et al., 2020) Denoising text inputs by deleting random tokens motivate autoencoders (AEs) to learn better latent representations. Our study compares 0%, 15%, and 40% deletion noising schemes.",
"cite_spans": [
{
"start": 183,
"end": 199,
"text": "Fu et al. (2019)",
"ref_id": "BIBREF4"
},
{
"start": 282,
"end": 304,
"text": "(Higgins et al., 2017)",
"ref_id": "BIBREF6"
},
{
"start": 476,
"end": 495,
"text": "(Shen et al., 2020)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "KL Weighting / Annealing",
"sec_num": "2.2.1"
},
{
"text": "KL-thresholding enforces a minimum \u03bb for each dimension of the KL term in the ELBO:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "KL thresholding (Kingma et al., 2016)",
"sec_num": "2.2.4"
},
{
"text": "L D KL = i max[\u03bb, D kl (q \u03c6 (z i |x)||p(z i ))] (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "KL thresholding (Kingma et al., 2016)",
"sec_num": "2.2.4"
},
{
"text": "where z i is a single dimension of z.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "KL thresholding (Kingma et al., 2016)",
"sec_num": "2.2.4"
},
{
"text": "Instead of using the last hidden state as encoder output, averaging or taking the maximum of all encoder hidden states results in a more diverse latent representation. We experiment with both meanand max-pooling schemes from the encoder.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder pooling (Long et al., 2019)",
"sec_num": "2.2.5"
},
{
"text": "We extend the T5 architecture (Raffel et al., 2020) into a VAE. We modify a popular pretrained T5 model (Wolf et al., 2020) that deviates minimally from the original Transformer (Figure 1 ).",
"cite_spans": [
{
"start": 30,
"end": 51,
"text": "(Raffel et al., 2020)",
"ref_id": "BIBREF18"
},
{
"start": 104,
"end": 123,
"text": "(Wolf et al., 2020)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [
{
"start": 178,
"end": 187,
"text": "(Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "Hidden states from all layers of T5's encoder q \u03c6 (z|x) are mean-or max-pooled into a vector h pooled \u2208 R H , where H is the encoder's hidden dimension.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "Assumed prior q(z)'s mean \u00b5 and log variance \u03c3 vectors of dimension L is obtained from h pooled :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "\u00b5 = h pooled W \u00b5 , log\u03c3 = h pooled W \u03c3 (3) where W \u00b5 , W \u03c3 \u2208 R L .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "As in a standard VAE, a stochastic latent vector z is sampled using the reparameterization trick to enable back-propagation through sampling:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "z = \u00b5 + \u03c3 , \u223c N (0, 1)",
"eq_num": "(4)"
}
],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "We pass z into the decoder p \u03b8 (x|z) as the only component of decoder's cross attention. Our Figure 1 : Transformer VAE architecture. A \"bottleneck\" step (W \u03c3 and W \u00b5 ) is placed between the encoder and the decoder of T5. Latent information from pooled encoder hidden states is captured in the bottleneck layer before being passed to the decoder. The network is optimized against regularization loss in the bottleneck and reconstruction loss at the decoder. method injects z into every layer of the decoder as in previous literature Fang et al., 2021) , but deviates in two important ways: first, we pass z as the sole key and value of encoder-decoder cross attention, instead of self-attention; second, we project z into the correct dimension (L \u00d7 A \u00d7 S, where L is the decoder layer count, A is the number of attention heads, and S is the embedding dimension per head) with a feed-forward network, instead of taking a copy of z to inject to each decoder layer.",
"cite_spans": [
{
"start": 533,
"end": 551,
"text": "Fang et al., 2021)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 93,
"end": 101,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "(K ca , V ca ) = (zW proj , zW proj )",
"eq_num": "(5)"
}
],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "where W proj \u2208 R L\u00d7A\u00d7S and K ca and V ca are key and value in decoder cross-attention.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model architecture",
"sec_num": "3"
},
{
"text": "During preliminary experiments, posterior collapse was observed in all training schemes without encoder warmup training. The decoder learns to ignore the initially noisy input signal from the encoder. Thus, we compose our finetuning method in two separate phases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Phase 1 -Encoder warmup: Weight of KL loss is set to zero, making our model's objective function similar to that of an AE. Different input denoising percentages, encoder pooling strategies, latent dimension sizes, and decoder freezing configurations are compared.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Phase 2 -Full finetuning: KL loss is reinstated and full VAE training are conducted. We compare different input denoising percentages, encoder pooling strategies, KL annealing schedules, and KL thresholds.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "We run our proposed two-phase finetuning training scheme on four standard VAE LM benchmark datasets: PTB (Marcus et al., 1993) , SNLI (Bowman et al., 2016), Yahoo (Yang et al., 2017) , and Yelp (Shen et al., 2017) .",
"cite_spans": [
{
"start": 105,
"end": 126,
"text": "(Marcus et al., 1993)",
"ref_id": "BIBREF16"
},
{
"start": 163,
"end": 182,
"text": "(Yang et al., 2017)",
"ref_id": "BIBREF24"
},
{
"start": 194,
"end": 213,
"text": "(Shen et al., 2017)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Following and Li et al. 2019, we perform intrinsic evaluation of our proposed Transformer VAE architecture. We report perplexity (PPL), KL-divergence between model posterior and assumed posterior (KL), and negative ELBO on the test set. To assess the quality of learned latent codes, we also report mutual information (MI) (Hoffman and Johnson, 2016) and the number of active units (AU) (Burda et al., 2016) . MI measures the dependence of latent codes to encoder input. AU measures the covariance between encoder input and latent codes.",
"cite_spans": [
{
"start": 387,
"end": 407,
"text": "(Burda et al., 2016)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Experimental hyperparameters such as specific annealing schedules and training epochs per phase are detailed in the appendix.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "We find that freezing the decoder and the memory projection layer W proj while training with an AE objective is crucial in learning meaningful encoder outputs. Denoising is important for datasets with longer inputs (Yahoo, Yelp), but not critical in datasets with shorter input lengths (PTB, SNLI). Mean-pooling encoder hidden states presents a trade-off between MI and AU. Max-pooling consistently learns more informative encoder representations. Changes in MI and AU during training is illustrated in Figure 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 503,
"end": 511,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Phase 1",
"sec_num": "5.1"
},
{
"text": "Latent dimensions of 64 and 128 were also tested. Increasing the latent dimension did not necessarily boost representational quality in terms",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phase 1",
"sec_num": "5.1"
},
{
"text": "PPL\u2193 KL -ELBO\u2193 MI\u2191 AU\u2191 Optimus (\u03bb = 0.5) 23.11 17.45 301.21 8.85 32 GPT-2 (Radford et al., 2019) 22.00 ----Encoder pretraining (\u03bb = 3) and . KLT denotes KL thresholding with \u03bb = 3. Our models are finetuned from a pretrained 6-layer T5, except the deep variant with 12 layers.",
"cite_spans": [
{
"start": 74,
"end": 96,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "of AU percentage. For latent dimensions of 32 and 64, 90% of latent dimension units were activated in best-performing models. For latent dimension of 128, around 60% of latent units were active. Another interesting observation is that KL divergence on the validation set, although not part of the AE training objective, plateaus after repeated training. We regard this phenomenon as the signal of convergence in terms of representation quality. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "We observe, as in previous literature, a trade-off between language modeling PPL and representation quality metrics (MI and AU). This trade-off is exacerbated when using KL thresholding. While KL thresholding does significantly increase latent representation capabilities, it is not in itself sufficient in preventing posterior collapse.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phase 2",
"sec_num": "5.2"
},
{
"text": "Denoising and encoder pooling configurations display the same characteristics as in Phase 1. No version of the experiment existed where cyclical annealing schedule was able to prevent posterior collapse, a result not in accordance with . Figure 3 illustrates the training progression of Phase 2. We also experimented with increasing model depth from 6 layers to 12 layers. Our proposed two-phase training scheme prevents posterior collapse for deeper models as well, resulting in higher performance in most metrics compared to 6-layer models. Results are reported in Table 1 . Note that lower PPL does not necessarily indicate better language modeling capabilities, since models with collapsed posterior display better PPL.",
"cite_spans": [],
"ref_spans": [
{
"start": 238,
"end": 246,
"text": "Figure 3",
"ref_id": "FIGREF1"
},
{
"start": 567,
"end": 574,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Phase 2",
"sec_num": "5.2"
},
{
"text": "Rows with KL above zero indicate successful aversion of posterior collapse. In the literature, no consensus yet exists on the optimal value of KL in training VAEs. Overall, we find that a denoising scheme between 0.15 and 0.4 in both phases, coupled with a low (0.5) KL threshold strikes a good balance between reconstruction and latent representation quality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phase 2",
"sec_num": "5.2"
},
{
"text": "This paper explores common methods in the literature for combatting posterior collapse, and the extent to which they help in teaching latent information to pretrained Transformer models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "Comprehensive experiments show that commonly employed posterior collapse mitigation techniques provide meaningful benefits in transforming existing language models into latent-aware architectures. Among the tested procedures, we find that 's two-step training, coupled with Shen et al. (2020) 's denoising through token deletion, was the most impactful in mitigating posterior collapse. However, language models obtained via only finetuning exhibit consistent trade-offs between their latent representation metrics (MI, AU) and language model metrics (PPL). Optimizing our model to be competitive with massively pretrained baselines in one of the two metrics results in the model falling behind in the other.",
"cite_spans": [
{
"start": 274,
"end": 292,
"text": "Shen et al. (2020)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "We also find that increasing training epochs further improves the impact of tested techniques, a result consistent with previous literature on largescale text VAE pretraining.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "From our experiments, we identify several questions to be answered by future research. The impact of homogenizing finetuning (as suggested in this paper) and original pretraining objectives on language model metrics has to be further explored. While the original T5 architecture was also pretrained with a self-supervised denoising scheme, the model employs mask tokens for denoising, contrary to simple token deletions suggested by this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "Our findings also highlight the need for an established heuristic to interpret the quality of latent representations learned by language models. The research community has yet to decide on the optimal value of KL-divergence between the assumed prior and the model posterior to target during text VAE training. Empirical guidelines to dictate even a vague threshold for the KL-divergence, below which we declare the occurrence of posterior collapse, will help both training and evaluation of latent-aware language models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "A Phase 2 results on PTB, Yelp, and SNLI Model PPL\u2193 KL -ELBO\u2193 MI\u2191 AU\u2191 Optimus (\u03bb = 0.5) 26.69 15.72 96.82 7.64 32 GPT-2 (Radford et al., 2019) 24.23 ----Encoder pretraining (\u03bb = 3) Model PPL\u2193 KL -ELBO\u2193 MI\u2191 AU\u2191 Optimus (\u03bb = 0.5) 22.79 15.09 344.10 9.13 32 GPT-2 (Radford et al., 2019) 23.40 ----Encoder pretraining (\u03bb = 3) (Li et al., Model PPL\u2193 KL -ELBO\u2193 MI\u2191 AU\u2191 Optimus (\u03bb = 0.5) 16.67 16.35 38.50 8.89 32 GPT-2 (Radford et al., 2019) 20.24 ----Encoder pretraining (\u03bb = 3) ",
"cite_spans": [
{
"start": 120,
"end": 142,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF17"
},
{
"start": 261,
"end": 283,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF17"
},
{
"start": 322,
"end": 333,
"text": "(Li et al.,",
"ref_id": null
},
{
"start": 413,
"end": 435,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "For all experiments we used a AdamW optimizer (Loshchilov and Hutter, 2019 ) with a starting learning rate of 1 \u00d7 10 \u22123 , \u03b2 1 = 0.9, \u03b2 2 = 0.999, and = 1 \u00d7 10 \u22123 . The linear KL annealing schedule we used was as follows:",
"cite_spans": [
{
"start": 46,
"end": 74,
"text": "(Loshchilov and Hutter, 2019",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "B Experimental details",
"sec_num": null
},
{
"text": "KL weight = current global step steps per epoch * 50",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Experimental details",
"sec_num": null
},
{
"text": "Our slower, linear KL annealing schedule of 0 to 1 over 50 epochs yielded better empirical results than the linear schdule used in Li et al. (2019) (0 to 1 over 10 epochs). We attribute this result to the small number of training samples in our experiments. We train for 5 epochs on Phase 1, and 3 epochs on Phase 2. While further training leads to increased MI and AU, we limit the number of epochs to confer to the spirit of this study, which is to learn latent representations with minimal training. The 5 epoch limit on Phase 1 was empirically determined as the point where encoder MI begins to plateau. Most experiments were conducted with z dimension of 32 for comparison with previous literature.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Experimental details",
"sec_num": null
},
{
"text": "https://github.com/seongminp/ transformers-into-vaes",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Generating sentences from a continuous space",
"authors": [
{
"first": "Samuel",
"middle": [],
"last": "Bowman",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Vilnis",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Dai",
"suffix": ""
},
{
"first": "Rafal",
"middle": [],
"last": "Jozefowicz",
"suffix": ""
},
{
"first": "Samy",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "10--21",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Samuel Bowman, Luke Vilnis, Oriol Vinyals, Andrew Dai, Rafal Jozefowicz, and Samy Bengio. 2016. Generating sentences from a continuous space. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 10-21.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Importance weighted autoencoders",
"authors": [
{
"first": "Yuri",
"middle": [],
"last": "Burda",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Roger",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Grosse",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
}
],
"year": 2016,
"venue": "ICLR (Poster)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuri Burda, Roger B Grosse, and Ruslan Salakhutdinov. 2016. Importance weighted autoencoders. In ICLR (Poster).",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Transformerbased conditional variational autoencoder for controllable story generation",
"authors": [
{
"first": "Le",
"middle": [],
"last": "Fang",
"suffix": ""
},
{
"first": "Tao",
"middle": [],
"last": "Zeng",
"suffix": ""
},
{
"first": "Chaochun",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Liefeng",
"middle": [],
"last": "Bo",
"suffix": ""
},
{
"first": "Wen",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Changyou",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.00828"
]
},
"num": null,
"urls": [],
"raw_text": "Le Fang, Tao Zeng, Chaochun Liu, Liefeng Bo, Wen Dong, and Changyou Chen. 2021. Transformer- based conditional variational autoencoder for controllable story generation. arXiv preprint arXiv:2101.00828.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Cyclical annealing schedule: A simple approach to mitigating kl vanishing",
"authors": [
{
"first": "Hao",
"middle": [],
"last": "Fu",
"suffix": ""
},
{
"first": "Chunyuan",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Xiaodong",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Asli",
"middle": [],
"last": "Celikyilmaz",
"suffix": ""
},
{
"first": "Lawrence",
"middle": [],
"last": "Carin",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "240--250",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hao Fu, Chunyuan Li, Xiaodong Liu, Jianfeng Gao, Asli Celikyilmaz, and Lawrence Carin. 2019. Cycli- cal annealing schedule: A simple approach to mit- igating kl vanishing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 240-250.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A probabilistic formulation of unsupervised text style transfer",
"authors": [
{
"first": "Junxian",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Xinyi",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
},
{
"first": "Taylor",
"middle": [],
"last": "Berg-Kirkpatrick",
"suffix": ""
}
],
"year": 2019,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. 2019. A probabilistic formulation of unsupervised text style transfer. In International Conference on Learning Representations.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "beta-vae: Learning basic visual concepts with a constrained variational framework",
"authors": [
{
"first": "Irina",
"middle": [],
"last": "Higgins",
"suffix": ""
},
{
"first": "Lo\u00efc",
"middle": [],
"last": "Matthey",
"suffix": ""
},
{
"first": "Arka",
"middle": [],
"last": "Pal",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Burgess",
"suffix": ""
},
{
"first": "Xavier",
"middle": [],
"last": "Glorot",
"suffix": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Botvinick",
"suffix": ""
},
{
"first": "Shakir",
"middle": [],
"last": "Mohamed",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Lerchner",
"suffix": ""
}
],
"year": 2017,
"venue": "5th International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Irina Higgins, Lo\u00efc Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. 2017. beta-vae: Learning basic visual concepts with a constrained variational framework. In 5th International Con- ference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Elbo surgery: yet another way to carve up the variational evidence lower bound",
"authors": [
{
"first": "D",
"middle": [],
"last": "Matthew",
"suffix": ""
},
{
"first": "Matthew J Johnson",
"middle": [],
"last": "Hoffman",
"suffix": ""
}
],
"year": 2016,
"venue": "Workshop in Advances in Approximate Bayesian Inference, NIPS",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew D Hoffman and Matthew J Johnson. 2016. Elbo surgery: yet another way to carve up the vari- ational evidence lower bound. In Workshop in Ad- vances in Approximate Bayesian Inference, NIPS, volume 1.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Semi-amortized variational autoencoders",
"authors": [
{
"first": "Yoon",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Wiseman",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Miller",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 35th International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "2678--2687",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoon Kim, Sam Wiseman, Andrew Miller, David Son- tag, and Alexander Rush. 2018. Semi-amortized variational autoencoders. In Proceedings of the 35th International Conference on Machine Learning, vol- ume 80 of Proceedings of Machine Learning Re- search, pages 2678-2687. PMLR.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Auto-Encoding Variational Bayes",
"authors": [
{
"first": "P",
"middle": [],
"last": "Diederik",
"suffix": ""
},
{
"first": "Max",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Welling",
"suffix": ""
}
],
"year": 2014,
"venue": "2nd International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diederik P. Kingma and Max Welling. 2014. Auto- Encoding Variational Bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Con- ference Track Proceedings.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Improved variational inference with inverse autoregressive flow",
"authors": [
{
"first": "P",
"middle": [],
"last": "Durk",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "Rafal",
"middle": [],
"last": "Salimans",
"suffix": ""
},
{
"first": "Xi",
"middle": [],
"last": "Jozefowicz",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Max",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Welling",
"suffix": ""
}
],
"year": 2016,
"venue": "Advances in Neural Information Processing Systems",
"volume": "29",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Durk P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. 2016. Improved variational inference with inverse autore- gressive flow. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A surprisingly effective fix for deep latent variable modeling of text",
"authors": [
{
"first": "Bohan",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Junxian",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
},
{
"first": "Taylor",
"middle": [],
"last": "Berg-Kirkpatrick",
"suffix": ""
},
{
"first": "Yiming",
"middle": [],
"last": "Yang",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "3603--3614",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1370"
]
},
"num": null,
"urls": [],
"raw_text": "Bohan Li, Junxian He, Graham Neubig, Taylor Berg- Kirkpatrick, and Yiming Yang. 2019. A surprisingly effective fix for deep latent variable modeling of text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), pages 3603- 3614, Hong Kong, China. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Optimus: Organizing sentences via pre-trained modeling of a latent space",
"authors": [
{
"first": "Chunyuan",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Xiang",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Yuan",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Baolin",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Xiujun",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Yizhe",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4678--4699",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiu- jun Li, Yizhe Zhang, and Jianfeng Gao. 2020. Opti- mus: Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 4678-4699.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A transformerbased variational autoencoder for sentence generation",
"authors": [
{
"first": "Danyang",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Gongshen",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "2019 International Joint Conference on Neural Networks (IJCNN)",
"volume": "",
"issue": "",
"pages": "1--7",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Danyang Liu and Gongshen Liu. 2019. A transformer- based variational autoencoder for sentence genera- tion. In 2019 International Joint Conference on Neu- ral Networks (IJCNN), pages 1-7. IEEE.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Preventing posterior collapse in sequence vaes with pooling. arXiv e-prints",
"authors": [
{
"first": "Teng",
"middle": [],
"last": "Long",
"suffix": ""
},
{
"first": "Yanshuai",
"middle": [],
"last": "Cao",
"suffix": ""
},
{
"first": "Jackie Chi Kit",
"middle": [],
"last": "Cheung",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Teng Long, Yanshuai Cao, and Jackie Chi Kit Cheung. 2019. Preventing posterior collapse in sequence vaes with pooling. arXiv e-prints, pages arXiv- 1911.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Decoupled weight decay regularization",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Loshchilov",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Hutter",
"suffix": ""
}
],
"year": 2019,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Con- ference on Learning Representations.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Building a large annotated corpus of English: The Penn Treebank",
"authors": [
{
"first": "Mitchell",
"middle": [
"P"
],
"last": "Marcus",
"suffix": ""
},
{
"first": "Beatrice",
"middle": [],
"last": "Santorini",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Ann"
],
"last": "Marcinkiewicz",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "313--330",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computa- tional Linguistics, 19(2):313-330.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Language models are unsupervised multitask learners",
"authors": [
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Rewon",
"middle": [],
"last": "Child",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Luan",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Amodei",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Exploring the limits of transfer learning with a unified text-totext transformer",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Roberts",
"suffix": ""
},
{
"first": "Katherine",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Sharan",
"middle": [],
"last": "Narang",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Matena",
"suffix": ""
},
{
"first": "Yanqi",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Peter",
"middle": [
"J"
],
"last": "Liu",
"suffix": ""
}
],
"year": 2020,
"venue": "Journal of Machine Learning Research",
"volume": "21",
"issue": "140",
"pages": "1--67",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Colin Raffel, Noam Shazeer, Adam Roberts, Kather- ine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to- text transformer. Journal of Machine Learning Re- search, 21(140):1-67.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Style transfer from non-parallel text by cross-alignment",
"authors": [
{
"first": "Tianxiao",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Tao",
"middle": [],
"last": "Lei",
"suffix": ""
},
{
"first": "Regina",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "Tommi",
"middle": [],
"last": "Jaakkola",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17",
"volume": "",
"issue": "",
"pages": "6833--6844",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. In Proceedings of the 31st Inter- national Conference on Neural Information Process- ing Systems, NIPS'17, page 6833-6844, Red Hook, NY, USA. Curran Associates Inc.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Educating text autoencoders: Latent representation guidance via denoising",
"authors": [
{
"first": "Tianxiao",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Jonas",
"middle": [],
"last": "Mueller",
"suffix": ""
},
{
"first": "Regina",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "Tommi",
"middle": [],
"last": "Jaakkola",
"suffix": ""
}
],
"year": 2020,
"venue": "International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "8719--8729",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tianxiao Shen, Jonas Mueller, Regina Barzilay, and Tommi Jaakkola. 2020. Educating text autoen- coders: Latent representation guidance via denois- ing. In International Conference on Machine Learn- ing, pages 8719-8729. PMLR.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Attention is all you need",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": ""
},
{
"first": "Illia",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Polosukhin",
"suffix": ""
}
],
"year": 2017,
"venue": "Advances in Neural Information Processing Systems",
"volume": "30",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Pro- cessing Systems, volume 30. Curran Associates, Inc.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Transformers: State-of-the-art natural language processing",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Wolf",
"suffix": ""
},
{
"first": "Lysandre",
"middle": [],
"last": "Debut",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Sanh",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Chaumond",
"suffix": ""
},
{
"first": "Clement",
"middle": [],
"last": "Delangue",
"suffix": ""
},
{
"first": "Anthony",
"middle": [],
"last": "Moi",
"suffix": ""
},
{
"first": "Pierric",
"middle": [],
"last": "Cistac",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Rault",
"suffix": ""
},
{
"first": "Remi",
"middle": [],
"last": "Louf",
"suffix": ""
},
{
"first": "Morgan",
"middle": [],
"last": "Funtowicz",
"suffix": ""
},
{
"first": "Joe",
"middle": [],
"last": "Davison",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Shleifer",
"suffix": ""
},
{
"first": "Clara",
"middle": [],
"last": "Patrick Von Platen",
"suffix": ""
},
{
"first": "Yacine",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Jernite",
"suffix": ""
},
{
"first": "Canwen",
"middle": [],
"last": "Plu",
"suffix": ""
},
{
"first": "Teven",
"middle": [
"Le"
],
"last": "Xu",
"suffix": ""
},
{
"first": "Sylvain",
"middle": [],
"last": "Scao",
"suffix": ""
},
{
"first": "Mariama",
"middle": [],
"last": "Gugger",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Drame",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
"volume": "",
"issue": "",
"pages": "38--45",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-demos.6"
]
},
"num": null,
"urls": [],
"raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, Remi Louf, Morgan Funtow- icz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Trans- formers: State-of-the-art natural language process- ing. In Proceedings of the 2020 Conference on Em- pirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Re-balancing variational autoencoder loss for molecule sequence generation",
"authors": [
{
"first": "Chaochao",
"middle": [],
"last": "Yan",
"suffix": ""
},
{
"first": "Sheng",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Jinyu",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Tingyang",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Junzhou",
"middle": [],
"last": "Huang",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 11th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics, BCB '20",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1145/3388440.3412458"
]
},
"num": null,
"urls": [],
"raw_text": "Chaochao Yan, Sheng Wang, Jinyu Yang, Tingyang Xu, and Junzhou Huang. 2020. Re-balancing variational autoencoder loss for molecule sequence generation. In Proceedings of the 11th ACM International Con- ference on Bioinformatics, Computational Biology and Health Informatics, BCB '20, New York, NY, USA. Association for Computing Machinery.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Improved variational autoencoders for text modeling using dilated convolutions",
"authors": [
{
"first": "Zichao",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zhiting",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "Taylor",
"middle": [],
"last": "Berg-Kirkpatrick",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 34th International Conference on Machine Learning",
"volume": "70",
"issue": "",
"pages": "3881--3890",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zichao Yang, Zhiting Hu, Ruslan Salakhutdinov, and Taylor Berg-Kirkpatrick. 2017. Improved varia- tional autoencoders for text modeling using dilated convolutions. In Proceedings of the 34th Interna- tional Conference on Machine Learning -Volume 70, ICML'17, page 3881-3890. JMLR.org.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Phase 1 training on Yahoo. Labels are in the form {pooling strategy}_{denoise percent-age}_{decoder frozen}.",
"type_str": "figure",
"num": null
},
"FIGREF1": {
"uris": null,
"text": "Phase 2 training on Yahoo. Labels are in the form {KL threshold}_{denoise percentage}. Encoder hidden states in plotted experiments were max-pooled.",
"type_str": "figure",
"num": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"text": "Phase 2 results on Yahoo. Due to space constraints, we report experimental results on other datasets in the appendix. Results on baselines are quoted from",
"html": null,
"num": null
},
"TABREF3": {
"type_str": "table",
"content": "<table/>",
"text": "Phase 2 results on PTB",
"html": null,
"num": null
},
"TABREF5": {
"type_str": "table",
"content": "<table/>",
"text": "Phase 2 results on Yelp.",
"html": null,
"num": null
},
"TABREF7": {
"type_str": "table",
"content": "<table/>",
"text": "Phase 2 results on SNLI.",
"html": null,
"num": null
}
}
}
} |