Spaces:
Running
Running
File size: 475,140 Bytes
14422b0 0a6b1f5 14422b0 ccd63d1 14422b0 0a6b1f5 14422b0 0a6b1f5 14422b0 ccd63d1 8c9e9ed 0a6b1f5 14422b0 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 18550fb 0a6b1f5 6d6a6aa 18550fb 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6d6a6aa 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 14422b0 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 ccd63d1 6b48067 14422b0 6b48067 14422b0 6b48067 14422b0 6b48067 14422b0 6b48067 14422b0 6b48067 14422b0 6b48067 14422b0 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 18550fb 6b48067 18550fb 6b48067 18550fb 6b48067 18550fb 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 18550fb 6b48067 18550fb 6b48067 0a6b1f5 18550fb 0a6b1f5 6b48067 18550fb 6b48067 18550fb 6b48067 18550fb 6b48067 0a6b1f5 ccd63d1 0a6b1f5 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 6b48067 ccd63d1 0a6b1f5 6b48067 0a6b1f5 6b48067 ccd63d1 6b48067 0a6b1f5 6af86f5 6b48067 6af86f5 6b48067 6af86f5 6b48067 0a6b1f5 6b48067 6af86f5 0a6b1f5 6b48067 6af86f5 6b48067 6af86f5 6b48067 6af86f5 6b48067 6af86f5 6b48067 6af86f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 ccd63d1 6b48067 0a6b1f5 6b48067 ccd63d1 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 8c9e9ed 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 ccd63d1 0a6b1f5 6b48067 0a6b1f5 6b48067 0a6b1f5 6b48067 14422b0 0a6b1f5 14422b0 0a6b1f5 14422b0 0a6b1f5 14422b0 0a6b1f5 14422b0 ccd63d1 8c9e9ed 0a6b1f5 14422b0 ccd63d1 14422b0 ccd63d1 14422b0 0a6b1f5 14422b0 0a6b1f5 14422b0 ccd63d1 14422b0 ccd63d1 14422b0 ccd63d1 14422b0 ccd63d1 14422b0 ccd63d1 14422b0 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 6af86f5 ccd63d1 14422b0 ccd63d1 0a6b1f5 14422b0 | 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 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225 6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645 6646 6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740 6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753 6754 6755 6756 6757 6758 6759 6760 6761 6762 6763 6764 6765 6766 6767 6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899 6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079 7080 7081 7082 7083 7084 7085 7086 7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111 7112 7113 7114 7115 7116 7117 7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134 7135 7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199 7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251 7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281 7282 7283 7284 7285 7286 7287 7288 7289 7290 7291 7292 7293 7294 7295 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306 7307 7308 7309 7310 7311 7312 7313 7314 7315 7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330 7331 7332 7333 7334 7335 7336 7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368 7369 7370 7371 7372 7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383 7384 7385 7386 7387 7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413 7414 7415 7416 7417 7418 7419 7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458 7459 7460 7461 7462 7463 7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489 7490 7491 7492 7493 7494 7495 7496 7497 7498 7499 7500 7501 7502 7503 7504 7505 7506 7507 7508 7509 7510 7511 7512 7513 7514 7515 7516 7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551 7552 7553 7554 7555 7556 7557 7558 7559 7560 7561 7562 7563 7564 7565 7566 7567 7568 7569 7570 7571 7572 7573 7574 7575 7576 7577 7578 7579 7580 7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604 7605 7606 7607 7608 7609 7610 7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621 7622 7623 7624 7625 7626 7627 7628 7629 7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655 7656 7657 7658 7659 7660 7661 7662 7663 7664 7665 7666 7667 7668 7669 7670 7671 7672 7673 7674 7675 7676 7677 7678 7679 7680 7681 7682 7683 7684 7685 7686 7687 7688 7689 7690 7691 7692 7693 7694 7695 7696 7697 7698 7699 7700 7701 7702 7703 7704 7705 7706 7707 7708 7709 7710 7711 7712 7713 7714 7715 7716 7717 7718 7719 7720 7721 7722 7723 7724 7725 7726 7727 7728 7729 7730 7731 7732 7733 7734 7735 7736 7737 7738 7739 7740 7741 7742 7743 7744 7745 7746 7747 7748 7749 7750 7751 7752 7753 7754 7755 7756 7757 7758 7759 7760 7761 7762 7763 7764 7765 7766 7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777 7778 7779 7780 7781 7782 7783 7784 7785 7786 7787 7788 7789 7790 7791 7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802 7803 7804 7805 7806 7807 7808 7809 7810 7811 7812 7813 7814 7815 7816 7817 7818 7819 7820 7821 7822 7823 7824 7825 7826 7827 7828 7829 7830 7831 7832 7833 7834 7835 7836 7837 7838 7839 7840 7841 7842 7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860 7861 7862 7863 7864 7865 7866 7867 7868 7869 7870 7871 7872 7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883 7884 7885 7886 7887 7888 7889 7890 7891 7892 7893 7894 7895 7896 7897 7898 7899 7900 7901 7902 7903 7904 7905 7906 7907 7908 7909 7910 7911 7912 7913 7914 7915 7916 7917 7918 7919 7920 7921 7922 7923 7924 7925 7926 7927 7928 7929 7930 7931 7932 7933 7934 7935 7936 7937 7938 7939 7940 7941 7942 7943 7944 7945 7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982 7983 7984 7985 7986 7987 7988 7989 7990 7991 7992 7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003 8004 8005 8006 8007 8008 8009 8010 8011 8012 8013 8014 8015 8016 8017 8018 8019 8020 8021 8022 8023 8024 8025 8026 8027 8028 8029 8030 8031 8032 8033 8034 8035 8036 8037 8038 8039 8040 8041 8042 8043 8044 8045 8046 8047 8048 8049 8050 8051 8052 8053 8054 8055 8056 8057 8058 8059 8060 8061 8062 8063 8064 8065 8066 8067 8068 8069 8070 8071 8072 8073 8074 8075 8076 8077 8078 8079 8080 8081 8082 8083 8084 8085 8086 8087 8088 8089 8090 8091 8092 8093 8094 8095 8096 8097 8098 8099 8100 8101 8102 8103 8104 8105 8106 8107 8108 8109 8110 8111 8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125 8126 8127 8128 8129 8130 8131 8132 8133 8134 8135 8136 8137 8138 8139 8140 8141 8142 8143 8144 8145 8146 8147 8148 8149 8150 8151 8152 8153 8154 8155 8156 8157 8158 8159 8160 8161 8162 8163 8164 8165 8166 8167 8168 8169 8170 8171 8172 8173 8174 8175 8176 8177 8178 8179 8180 8181 8182 8183 8184 8185 8186 8187 8188 8189 8190 8191 8192 8193 8194 8195 8196 8197 8198 8199 8200 8201 8202 8203 8204 8205 8206 8207 8208 8209 8210 8211 8212 8213 8214 8215 8216 8217 8218 8219 8220 8221 8222 8223 8224 8225 8226 8227 8228 8229 8230 8231 8232 8233 8234 8235 8236 8237 8238 8239 8240 8241 8242 8243 8244 8245 8246 8247 8248 8249 8250 8251 8252 8253 8254 8255 8256 8257 8258 8259 8260 8261 8262 8263 8264 8265 8266 8267 8268 8269 8270 8271 8272 8273 8274 8275 8276 8277 8278 8279 8280 8281 8282 8283 8284 8285 8286 8287 8288 8289 8290 8291 8292 8293 8294 8295 8296 8297 8298 8299 8300 8301 8302 8303 8304 8305 8306 8307 8308 8309 8310 8311 8312 8313 8314 8315 8316 8317 8318 8319 8320 8321 8322 8323 8324 8325 8326 8327 8328 8329 8330 8331 8332 8333 8334 8335 8336 8337 8338 8339 8340 8341 8342 8343 8344 8345 8346 8347 8348 8349 8350 8351 8352 8353 8354 8355 8356 8357 8358 8359 8360 8361 8362 8363 8364 8365 8366 8367 8368 8369 8370 8371 8372 8373 8374 8375 8376 8377 8378 8379 8380 8381 8382 8383 8384 8385 8386 8387 8388 8389 8390 8391 8392 8393 8394 8395 8396 8397 8398 8399 8400 8401 8402 8403 8404 8405 8406 8407 8408 8409 8410 8411 8412 8413 8414 8415 8416 8417 8418 8419 8420 8421 8422 8423 8424 8425 8426 8427 8428 8429 8430 8431 8432 8433 8434 8435 8436 8437 8438 8439 8440 8441 8442 8443 8444 8445 8446 8447 8448 8449 8450 8451 8452 8453 8454 8455 8456 8457 8458 8459 8460 8461 8462 8463 8464 8465 8466 8467 8468 8469 8470 8471 8472 8473 8474 8475 8476 8477 8478 8479 8480 8481 8482 8483 8484 8485 8486 8487 8488 8489 8490 8491 8492 8493 8494 8495 8496 8497 8498 8499 8500 8501 8502 8503 8504 8505 8506 8507 8508 8509 8510 8511 8512 8513 8514 8515 8516 8517 8518 8519 8520 8521 8522 8523 8524 8525 8526 8527 8528 8529 8530 8531 8532 8533 8534 8535 8536 8537 8538 8539 8540 8541 8542 8543 8544 8545 8546 8547 8548 8549 8550 8551 8552 8553 8554 8555 8556 8557 8558 8559 8560 8561 8562 8563 8564 8565 8566 8567 8568 8569 8570 8571 8572 8573 8574 8575 8576 8577 8578 8579 8580 8581 8582 8583 8584 8585 8586 8587 8588 8589 8590 8591 8592 8593 8594 8595 8596 8597 8598 8599 8600 8601 8602 8603 8604 8605 8606 8607 8608 8609 8610 8611 8612 8613 8614 8615 8616 8617 8618 8619 8620 8621 8622 8623 8624 8625 8626 8627 8628 8629 8630 8631 8632 8633 8634 8635 8636 8637 8638 8639 8640 8641 8642 8643 8644 8645 8646 8647 8648 8649 8650 8651 8652 8653 8654 8655 8656 8657 8658 | <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Complete Deep Learning & Computer Vision Curriculum</title>
<!-- Shared Design System -->
<link rel="stylesheet" href="../shared/css/design-system.css">
<link rel="stylesheet" href="../shared/css/components.css">
<!-- MathJax 3.x -->
<script>
MathJax = { tex: { inlineMath: [['$', '$'], ['\\(', '\\)']] } };
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<style>
* {
margin: 0;
padding: 0;
box-sizing: border-box;
}
:root {
--bg: #0f1419;
--surface: #1a1f2e;
--text: #e4e6eb;
--text-dim: #b0b7c3;
--cyan: #00d4ff;
--orange: #ff6b35;
--green: #00ff88;
--yellow: #ffa500;
}
body {
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;
background: var(--bg);
color: var(--text);
line-height: 1.6;
overflow-x: hidden;
}
.container {
max-width: 1400px;
margin: 0 auto;
padding: 20px;
}
header {
text-align: center;
margin-bottom: 40px;
padding: 30px 0;
border-bottom: 2px solid var(--cyan);
}
h1 {
font-size: 2.5em;
background: linear-gradient(135deg, var(--cyan), var(--orange));
background-clip: text;
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
margin-bottom: 10px;
}
.subtitle {
color: var(--text-dim);
font-size: 1.1em;
}
.dashboard {
display: none;
}
.dashboard.active {
display: block;
}
.grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(280px, 1fr));
gap: 25px;
margin: 40px 0;
}
.card {
background: linear-gradient(135deg, rgba(0, 212, 255, 0.1), rgba(255, 107, 53, 0.1));
border: 2px solid var(--cyan);
border-radius: 12px;
padding: 30px;
cursor: pointer;
transition: all 0.3s ease;
text-align: center;
}
.card:hover {
transform: translateY(-5px);
box-shadow: 0 10px 30px rgba(0, 212, 255, 0.2);
border-color: var(--orange);
}
.card-icon {
font-size: 3em;
margin-bottom: 15px;
}
.card h3 {
color: var(--cyan);
font-size: 1.5em;
margin-bottom: 10px;
}
.card p {
color: var(--text-dim);
font-size: 0.95em;
}
.category-label {
display: inline-block;
margin-top: 10px;
padding: 5px 12px;
background: rgba(0, 212, 255, 0.2);
border-radius: 20px;
font-size: 0.85em;
color: var(--green);
}
.module {
display: none;
}
.module.active {
display: block;
animation: fadeIn 0.3s ease;
}
@keyframes fadeIn {
from {
opacity: 0;
}
to {
opacity: 1;
}
}
.btn-back {
padding: 10px 20px;
background: var(--orange);
color: var(--bg);
border: none;
border-radius: 6px;
cursor: pointer;
font-weight: 600;
margin-bottom: 25px;
transition: all 0.3s ease;
}
.btn-back:hover {
background: var(--cyan);
}
.tabs {
display: flex;
gap: 10px;
margin-bottom: 30px;
flex-wrap: wrap;
justify-content: center;
border-bottom: 1px solid rgba(0, 212, 255, 0.2);
padding-bottom: 15px;
overflow-x: auto;
}
.tab-btn {
padding: 10px 20px;
background: var(--surface);
color: var(--text);
border: 2px solid transparent;
border-radius: 6px;
cursor: pointer;
font-size: 0.95em;
transition: all 0.3s ease;
font-weight: 500;
white-space: nowrap;
}
.tab-btn:hover {
background: rgba(0, 212, 255, 0.1);
border-color: var(--cyan);
}
.tab-btn.active {
background: var(--cyan);
color: var(--bg);
border-color: var(--cyan);
}
.tab {
display: none;
}
.tab.active {
display: block;
animation: fadeIn 0.3s ease;
}
.section {
background: var(--surface);
border: 1px solid rgba(0, 212, 255, 0.2);
border-radius: 10px;
padding: 30px;
margin-bottom: 25px;
transition: all 0.3s ease;
}
.section:hover {
border-color: var(--cyan);
box-shadow: 0 0 20px rgba(0, 212, 255, 0.1);
}
h2 {
color: var(--cyan);
font-size: 1.8em;
margin-bottom: 15px;
}
h3 {
color: var(--orange);
font-size: 1.3em;
margin-top: 20px;
margin-bottom: 12px;
}
h4 {
color: var(--green);
font-size: 1.1em;
margin-top: 15px;
margin-bottom: 10px;
}
p {
margin-bottom: 15px;
line-height: 1.8;
}
ul {
margin-left: 20px;
margin-bottom: 15px;
}
ul li {
margin-bottom: 8px;
}
.info-box {
background: linear-gradient(135deg, rgba(0, 212, 255, 0.1), rgba(255, 107, 53, 0.1));
border: 1px solid var(--cyan);
border-radius: 8px;
padding: 20px;
margin: 20px 0;
}
.box-title {
color: var(--orange);
font-weight: 700;
margin-bottom: 10px;
font-size: 1.1em;
}
.box-content {
color: var(--text-dim);
line-height: 1.7;
}
.formula {
background: rgba(0, 212, 255, 0.1);
border: 1px solid var(--cyan);
border-radius: 8px;
padding: 20px;
margin: 20px 0;
font-family: 'Courier New', monospace;
overflow-x: auto;
line-height: 1.8;
color: var(--cyan);
}
.code-block {
background: #0d1117;
border: 1px solid #30363d;
border-radius: 8px;
padding: 20px;
margin: 20px 0;
overflow-x: auto;
font-family: 'Fira Code', 'Consolas', monospace;
font-size: 0.88em;
line-height: 1.7;
white-space: pre-wrap;
color: #e6edf3;
}
.code-block .comment {
color: #8b949e;
}
.code-block .keyword {
color: #ff7b72;
}
.code-block .string {
color: #a5d6ff;
}
.code-block .function {
color: #d2a8ff;
}
.code-block .number {
color: #79c0ff;
}
.code-block .class-name {
color: #ffa657;
}
.code-block .builtin {
color: #79c0ff;
}
.code-title {
display: block;
background: #161b22;
color: #8b949e;
padding: 8px 16px;
border-radius: 8px 8px 0 0;
border: 1px solid #30363d;
border-bottom: none;
font-size: 0.85em;
font-family: 'Fira Code', monospace;
margin-top: 20px;
}
.code-title+.code-block {
margin-top: 0;
border-radius: 0 0 8px 8px;
}
.callout {
border-left: 4px solid;
padding: 15px;
margin: 20px 0;
border-radius: 6px;
}
.callout.tip {
border-left-color: var(--green);
background: rgba(0, 255, 136, 0.05);
}
.callout.warning {
border-left-color: var(--yellow);
background: rgba(255, 165, 0, 0.05);
}
.callout.insight {
border-left-color: var(--cyan);
background: rgba(0, 212, 255, 0.05);
}
.callout-title {
font-weight: 700;
margin-bottom: 8px;
}
.list-item {
display: flex;
gap: 12px;
margin: 12px 0;
padding: 12px;
background: rgba(0, 212, 255, 0.05);
border-left: 3px solid var(--cyan);
border-radius: 4px;
}
.list-num {
color: var(--orange);
font-weight: 700;
min-width: 30px;
}
table {
width: 100%;
border-collapse: collapse;
margin: 20px 0;
}
th,
td {
padding: 12px;
text-align: left;
border: 1px solid rgba(0, 212, 255, 0.2);
}
th {
background: rgba(0, 212, 255, 0.1);
color: var(--cyan);
font-weight: 700;
}
.viz-container {
background: rgba(0, 212, 255, 0.02);
border: 1px solid rgba(0, 212, 255, 0.2);
border-radius: 8px;
padding: 20px;
margin: 20px 0;
display: flex;
justify-content: center;
overflow-x: auto;
}
.viz-controls {
display: flex;
gap: 10px;
margin-top: 20px;
justify-content: center;
flex-wrap: wrap;
}
.btn-viz {
padding: 10px 20px;
background: var(--cyan);
color: var(--bg);
border: none;
border-radius: 6px;
font-weight: 600;
cursor: pointer;
font-size: 0.95em;
transition: all 0.3s ease;
}
.btn-viz:hover {
background: var(--orange);
transform: scale(1.05);
}
canvas {
max-width: 100%;
height: auto;
}
@media (max-width: 768px) {
h1 {
font-size: 1.8em;
}
.tabs {
flex-direction: column;
}
.tab-btn {
width: 100%;
}
.grid {
grid-template-columns: 1fr;
}
canvas {
width: 100% !important;
height: auto !important;
}
}
</style>
</head>
<body>
<div class="container">
<!-- MAIN DASHBOARD -->
<div id="dashboard" class="dashboard active">
<header>
<h1>๐ง Complete Deep Learning & Computer Vision</h1>
<p class="subtitle">Comprehensive Curriculum | Foundations to Advanced Applications</p>
</header>
<div style="text-align: center; margin-bottom: 40px;">
<p style="color: var(--text-dim); font-size: 1.1em;">
Master all aspects of deep learning and computer vision. 25+ modules covering neural networks, CNNs,
object detection, GANs, and more.
</p>
</div>
<div class="grid" id="modulesGrid"></div>
</div>
<!-- MODULES CONTAINER -->
<div id="modulesContainer"></div>
</div>
<script>
const modules = [
// Module 1: Deep Learning Foundations
{
id: "nn-basics",
title: "Introduction to Neural Networks",
icon: "๐งฌ",
category: "Foundations",
color: "#0088ff",
description: "Biological vs. Artificial neurons and network architecture"
},
{
id: "perceptron",
title: "The Perceptron",
icon: "โ๏ธ",
category: "Foundations",
color: "#0088ff",
description: "Single layer networks and their limitations"
},
{
id: "mlp",
title: "Multi-Layer Perceptron (MLP)",
icon: "๐๏ธ",
category: "Foundations",
color: "#0088ff",
description: "Hidden layers and deep architectures"
},
{
id: "activation",
title: "Activation Functions",
icon: "โก",
category: "Foundations",
color: "#0088ff",
description: "Sigmoid, ReLU, Tanh, Leaky ReLU, ELU, Softmax"
},
{
id: "weight-init",
title: "Weight Initialization",
icon: "๐ฏ",
category: "Foundations",
color: "#0088ff",
description: "Xavier, He, Random initialization strategies"
},
{
id: "loss",
title: "Loss Functions",
icon: "๐",
category: "Foundations",
color: "#0088ff",
description: "MSE, Binary Cross-Entropy, Categorical Cross-Entropy"
},
{
id: "optimizers",
title: "Optimizers",
icon: "๐ฏ",
category: "Training",
color: "#00ff00",
description: "SGD, Momentum, Adam, Adagrad, RMSprop"
},
{
id: "backprop",
title: "Forward & Backpropagation",
icon: "โฌ
๏ธ",
category: "Training",
color: "#00ff00",
description: "Chain rule and gradient computation"
},
{
id: "regularization",
title: "Regularization",
icon: "๐ก๏ธ",
category: "Training",
color: "#00ff00",
description: "L1/L2, Dropout, Early Stopping, Batch Norm"
},
{
id: "batch-norm",
title: "Batch Normalization",
icon: "โ๏ธ",
category: "Training",
color: "#00ff00",
description: "Stabilizing and speeding up training"
},
// Module 2: Computer Vision Fundamentals
{
id: "cv-intro",
title: "CV Fundamentals",
icon: "๐๏ธ",
category: "Computer Vision",
color: "#ff6b35",
description: "Why ANNs fail with images, parameter explosion"
},
{
id: "conv-layer",
title: "Convolutional Layers",
icon: "๐ผ๏ธ",
category: "Computer Vision",
color: "#ff6b35",
description: "Kernels, filters, feature maps, stride, padding"
},
{
id: "pooling",
title: "Pooling Layers",
icon: "๐ฆ",
category: "Computer Vision",
color: "#ff6b35",
description: "Max pooling, average pooling, spatial reduction"
},
{
id: "cnn-basics",
title: "CNN Architecture",
icon: "๐๏ธ",
category: "Computer Vision",
color: "#ff6b35",
description: "Combining conv, pooling, and fully connected layers"
},
{
id: "viz-filters",
title: "Visualizing CNNs",
icon: "๐",
category: "Computer Vision",
color: "#ff6b35",
description: "What filters learn: edges โ shapes โ objects"
},
// Module 3: Advanced CNN Architectures
{
id: "lenet",
title: "LeNet-5",
icon: "๐ข",
category: "CNN Architectures",
color: "#ff00ff",
description: "Classic digit recognizer (MNIST)"
},
{
id: "alexnet",
title: "AlexNet",
icon: "๐",
category: "CNN Architectures",
color: "#ff00ff",
description: "The breakthrough in deep computer vision (2012)"
},
{
id: "vgg",
title: "VGGNet",
icon: "๐",
category: "CNN Architectures",
color: "#ff00ff",
description: "VGG-16/19: Deep networks with small filters"
},
{
id: "resnet",
title: "ResNet",
icon: "๐",
category: "CNN Architectures",
color: "#ff00ff",
description: "Skip connections, solving vanishing gradients"
},
{
id: "inception",
title: "InceptionNet (GoogLeNet)",
icon: "๐ฏ",
category: "CNN Architectures",
color: "#ff00ff",
description: "1x1 convolutions, multi-scale feature extraction"
},
{
id: "mobilenet",
title: "MobileNet",
icon: "๐ฑ",
category: "CNN Architectures",
color: "#ff00ff",
description: "Depth-wise separable convolutions for efficiency"
},
{
id: "transfer-learning",
title: "Transfer Learning",
icon: "๐",
category: "CNN Architectures",
color: "#ff00ff",
description: "Fine-tuning and leveraging pre-trained models"
},
// Module 4: Object Detection & Segmentation
{
id: "localization",
title: "Object Localization",
icon: "๐",
category: "Detection",
color: "#00ff00",
description: "Bounding boxes and classification together"
},
{
id: "rcnn",
title: "R-CNN Family",
icon: "๐ฏ",
category: "Detection",
color: "#00ff00",
description: "R-CNN, Fast R-CNN, Faster R-CNN"
},
{
id: "yolo",
title: "YOLO",
icon: "โก",
category: "Detection",
color: "#00ff00",
description: "Real-time object detection (v3, v5, v8)"
},
{
id: "ssd",
title: "SSD",
icon: "๐",
category: "Detection",
color: "#00ff00",
description: "Single Shot MultiBox Detector"
},
{
id: "semantic-seg",
title: "Semantic Segmentation",
icon: "๐๏ธ",
category: "Segmentation",
color: "#00ff00",
description: "Pixel-level classification (U-Net)"
},
{
id: "instance-seg",
title: "Instance Segmentation",
icon: "๐ฅ",
category: "Segmentation",
color: "#00ff00",
description: "Mask R-CNN and separate object instances"
},
{
id: "face-recog",
title: "Face Recognition",
icon: "๐ค",
category: "Segmentation",
color: "#00ff00",
description: "Siamese networks and triplet loss"
},
// Module 5: Generative Models
{
id: "autoencoders",
title: "Autoencoders",
icon: "๐",
category: "Generative",
color: "#ffaa00",
description: "Encoder-decoder, latent space, denoising"
},
{
id: "gans",
title: "GANs (Generative Adversarial Networks)",
icon: "๐ฎ",
category: "Generative",
color: "#ffaa00",
description: "Generator vs. Discriminator, DCGAN"
},
{
id: "diffusion",
title: "Diffusion Models",
icon: "๐",
category: "Generative",
color: "#ffaa00",
description: "Foundation of Stable Diffusion and DALL-E"
},
// Additional Advanced Topics
{
id: "rnn",
title: "RNNs & LSTMs",
icon: "๐",
category: "Sequence",
color: "#ff6b35",
description: "Recurrent networks for sequential data"
},
{
id: "transformers",
title: "Transformers",
icon: "๐",
category: "Sequence",
color: "#ff6b35",
description: "\"Attention Is All You Need\" - Complete paper breakdown with math"
},
{
id: "bert",
title: "BERT & NLP Transformers",
icon: "๐",
category: "NLP",
color: "#ff6b35",
description: "Bidirectional transformers for language"
},
{
id: "gpt",
title: "GPT & Language Models",
icon: "๐ฌ",
category: "NLP",
color: "#ff6b35",
description: "Autoregressive models and text generation"
},
{
id: "vit",
title: "Vision Transformers (ViT)",
icon: "๐จ",
category: "Vision",
color: "#ff6b35",
description: "Transformers applied to image data"
},
{
id: "gnn",
title: "Graph Neural Networks",
icon: "๐ธ๏ธ",
category: "Advanced",
color: "#9900ff",
description: "Deep learning on non-Euclidean graph data"
},
{
id: "seq2seq",
title: "Seq2Seq & Attention",
icon: "โก๏ธ",
category: "NLP",
color: "#ff6b35",
description: "Encoder-Decoder models and Attention mechanics"
},
{
id: "research-papers",
title: "Research Library",
icon: "๐",
category: "Advanced",
color: "#9900ff",
description: "Curated collection of seminal deep learning papers"
},
// Module 7: GenAI & LLM Engineering
{
id: "vector-db",
title: "Vector Databases",
icon: "๐งฒ",
category: "GenAI",
color: "#00c9a7",
description: "Embeddings, similarity search, FAISS, Pinecone, ChromaDB"
},
{
id: "rag",
title: "RAG Pipelines",
icon: "๐",
category: "GenAI",
color: "#00c9a7",
description: "Retrieval-Augmented Generation for grounded AI"
},
{
id: "advanced-llm",
title: "Fine-Tuning & Quantization",
icon: "โ๏ธ",
category: "GenAI",
color: "#00c9a7",
description: "LoRA, QLoRA, PEFT, GGUF, and deployment strategies"
}
];
// Comprehensive content for all modules
const MODULE_CONTENT = {
"nn-basics": {
overview: `
<p>A neural network is a computing system <strong>inspired by biological brains</strong> โ but it's not a brain simulation. It's a mathematical function that learns patterns from data by adjusting millions of numbers called <strong>weights</strong>.</p>
<div class="callout insight">
<div class="callout-title">๐ง The Core Intuition</div>
Think of a neural network like a factory assembly line. Raw materials (data) enter one end. At each station (layer), workers (neurons) examine the material, make a small decision, and pass the result forward. By the end, the factory produces a finished product (prediction). During training, a quality inspector (loss function) checks the output and sends feedback backward through the line so each worker can improve.
</div>
<h3>How Does a Single Neuron Work?</h3>
<p>A neuron does three things:</p>
<div class="list-item">
<div class="list-num">1</div>
<div><strong>Weighted Sum:</strong> Multiply each input by a weight and add them up: $$z = w_1 x_1 + w_2 x_2 + ... + w_n x_n + b$$</div>
</div>
<div class="list-item">
<div class="list-num">2</div>
<div><strong>Bias:</strong> Add a bias term $b$ โ this shifts the decision boundary (like the y-intercept in y = mx + b)</div>
</div>
<div class="list-item">
<div class="list-num">3</div>
<div><strong>Activation:</strong> Pass the result through a non-linear function: $$a = \\sigma(z)$$</div>
</div>
<h3>The Training Process</h3>
<p>Training a neural network follows a simple loop repeated thousands of times:</p>
<div class="info-box">
<div class="box-title">๐ The Training Loop</div>
<div class="box-content">
<strong>1. Forward Pass:</strong> Feed data through the network, get a prediction<br>
<strong>2. Compute Loss:</strong> Measure how wrong the prediction is (e.g., MSE, Cross-Entropy)<br>
<strong>3. Backward Pass:</strong> Calculate how each weight contributed to the error (gradients via chain rule)<br>
<strong>4. Update Weights:</strong> Nudge each weight slightly to reduce the error: $$w_{new} = w_{old} - \\eta \\cdot \\frac{\\partial L}{\\partial w}$$
<strong>5. Repeat</strong> until the loss is small enough
</div>
</div>
<h3>Types of Layers</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; text-align: left; color: var(--cyan);">Layer Type</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Purpose</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Example</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Dense/Linear</td><td>Every neuron connects to every input</td><td>nn.Linear(784, 128)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Convolutional</td><td>Detect spatial patterns (edges, textures)</td><td>nn.Conv2d(3, 32, 3)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Recurrent (LSTM/GRU)</td><td>Process sequences (text, time series)</td><td>nn.LSTM(128, 256)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Attention</td><td>Focus on relevant parts of input</td><td>nn.MultiheadAttention()</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Dropout</td><td>Regularization (randomly zero neurons)</td><td>nn.Dropout(0.5)</td></tr>
<tr><td style="padding: 8px;">BatchNorm</td><td>Stabilize and speed up training</td><td>nn.BatchNorm1d(128)</td></tr>
</table>
`,
concepts: `
<h3>Core Components</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Neurons (Nodes):</strong> Basic computational units that receive inputs, apply weights, add bias, and apply activation function</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Layers:</strong> Input layer (receives data), Hidden layers (feature extraction), Output layer (predictions)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Weights:</strong> Parameters learned during training that determine connection strength</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Bias:</strong> Allows shifting the activation function for better fitting</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Activation Function:</strong> Introduces non-linearity (ReLU, Sigmoid, Tanh)</div>
</div>
`,
applications: `
<h3>Real-World Applications</h3>
<div class="info-box">
<div class="box-title">๐ฅ Healthcare</div>
<div class="box-content">Disease diagnosis, medical image analysis, drug discovery, patient risk prediction</div>
</div>
<div class="info-box">
<div class="box-title">๐ฐ Finance</div>
<div class="box-content">Fraud detection, algorithmic trading, credit scoring, portfolio optimization</div>
</div>
<div class="info-box">
<div class="box-title">๐ E-commerce</div>
<div class="box-content">Recommendation systems, demand forecasting, customer segmentation, price optimization</div>
</div>
`,
math: `
<h3>The Fundamental Equations of a Neuron</h3>
<p>A single neuron performs a weighted sum followed by an activation function. This is the atomic building block of all neural networks.</p>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>z = ฮฃ(wแตขxแตข) + b = wโxโ + wโxโ + ... + wโxโ + b</strong><br>
<strong>a = ฯ(z)</strong>
</div>
<h4>Step-by-Step: Single Neuron Forward Pass</h4>
<div class="list-item">
<div class="list-num">01</div>
<div>
<strong>Weighted Sum (Linear):</strong><br>
z = wแตx + b = ฮฃแตข wแตขxแตข + b<br>
<span class="formula-caption">This is a dot product plus bias - pure linear algebra</span>
</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div>
<strong>Activation (Non-Linear):</strong><br>
a = ฯ(z) where ฯ can be ReLU, Sigmoid, Tanh, etc.<br>
<span class="formula-caption">This introduces non-linearity, enabling complex functions</span>
</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Manual Forward Pass</div>
<strong>Inputs:</strong> x = [2, 3], <strong>Weights:</strong> w = [0.5, -0.3], <strong>Bias:</strong> b = 0.1<br><br>
<strong>Step 1 - Weighted Sum:</strong><br>
z = (0.5 ร 2) + (-0.3 ร 3) + 0.1<br>
z = 1.0 - 0.9 + 0.1 = <strong>0.2</strong><br><br>
<strong>Step 2 - ReLU Activation:</strong><br>
a = max(0, 0.2) = <strong>0.2</strong><br><br>
<strong>Step 2 (alt) - Sigmoid Activation:</strong><br>
a = 1 / (1 + eโปโฐยทยฒ) = 1 / 1.819 โ <strong>0.55</strong>
</div>
<h3>Network Layer in Matrix Form</h3>
<p>For a layer with n inputs and m neurons, we use matrices for efficiency:</p>
<div class="formula">
<strong>Z = WX + b</strong><br>
<strong>A = ฯ(Z)</strong><br><br>
Where:<br>
โข W โ โแตหฃโฟ (weight matrix: m neurons, n inputs)<br>
โข X โ โโฟหฃยน (input vector)<br>
โข b โ โแตหฃยน (bias vector)<br>
โข Z โ โแตหฃยน (pre-activation)<br>
โข A โ โแตหฃยน (activation output)
</div>
<h3>Parameter Count Formula</h3>
<div class="formula">
For a layer: n_in โ n_out<br>
<strong>Parameters = n_in ร n_out + n_out</strong><br>
(weights) + (biases)<br><br>
Example: Layer 784 โ 128<br>
Params = 784 ร 128 + 128 = 100,480
</div>
`,
code: `
<h3>Build a Neuron from Scratch (NumPy)</h3>
<p>Let's implement the fundamental unit of a neural network โ a single neuron โ using only NumPy. This helps you understand what happens <strong>under the hood</strong> before using frameworks like PyTorch.</p>
<span class="code-title">๐ neuron_numpy.py</span><div class="code-block"><span class="keyword">import</span> numpy <span class="keyword">as</span> np
<span class="comment"># --- A Single Neuron from Scratch ---</span>
<span class="keyword">def</span> <span class="function">sigmoid</span>(z):
<span class="string">"""Activation function: squashes output to (0, 1)"""</span>
<span class="keyword">return</span> <span class="number">1</span> / (<span class="number">1</span> + np.exp(-z))
<span class="comment"># Input features (e.g., hours studied, hours slept)</span>
X = np.array([<span class="number">6.0</span>, <span class="number">8.0</span>])
<span class="comment"># Learnable parameters</span>
weights = np.array([<span class="number">0.5</span>, <span class="number">-0.3</span>])
bias = <span class="number">0.1</span>
<span class="comment"># Forward pass: z = wยทx + b</span>
z = np.dot(weights, X) + bias <span class="comment"># 0.5*6 + (-0.3)*8 + 0.1 = 0.7</span>
output = sigmoid(z) <span class="comment"># ฯ(0.7) โ 0.668</span>
<span class="keyword">print</span>(f<span class="string">"Linear output z = {z:.3f}"</span>)
<span class="keyword">print</span>(f<span class="string">"Activated output a = {output:.3f}"</span>)</div>
<h3>The Same Neuron in PyTorch</h3>
<p>PyTorch wraps this into <code>nn.Linear</code>. One line replaces all the manual math above โ but the computation is identical.</p>
<span class="code-title">๐ neuron_pytorch.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="comment"># A single neuron: 2 inputs โ 1 output</span>
neuron = nn.Linear(in_features=<span class="number">2</span>, out_features=<span class="number">1</span>)
<span class="comment"># Input tensor</span>
X = torch.tensor([[<span class="number">6.0</span>, <span class="number">8.0</span>]])
<span class="comment"># Forward pass</span>
z = neuron(X) <span class="comment"># Linear: z = Wx + b</span>
a = torch.sigmoid(z) <span class="comment"># Activation</span>
<span class="keyword">print</span>(f<span class="string">"Output: {a.item():.4f}"</span>)
<span class="comment"># Check learnable parameters</span>
<span class="keyword">print</span>(f<span class="string">"Weights: {neuron.weight.data}"</span>)
<span class="keyword">print</span>(f<span class="string">"Bias: {neuron.bias.data}"</span>)
<span class="keyword">print</span>(f<span class="string">"Total params: {sum(p.numel() for p in neuron.parameters())}"</span>) <span class="comment"># 3</span></div>
<h3>Build a Simple Neural Network</h3>
<span class="code-title">๐ simple_network.py</span><div class="code-block"><span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="comment"># 3-layer network: 784 โ 128 โ 64 โ 10</span>
model = nn.Sequential(
nn.Linear(<span class="number">784</span>, <span class="number">128</span>), <span class="comment"># Hidden layer 1</span>
nn.ReLU(),
nn.Linear(<span class="number">128</span>, <span class="number">64</span>), <span class="comment"># Hidden layer 2</span>
nn.ReLU(),
nn.Linear(<span class="number">64</span>, <span class="number">10</span>), <span class="comment"># Output layer (10 classes)</span>
)
<span class="comment"># Count parameters</span>
total = <span class="builtin">sum</span>(p.numel() <span class="keyword">for</span> p <span class="keyword">in</span> model.parameters())
<span class="keyword">print</span>(f<span class="string">"Total parameters: {total:,}"</span>) <span class="comment"># 109,386</span>
<span class="comment"># Forward pass with random MNIST-like input</span>
X = torch.randn(<span class="number">1</span>, <span class="number">784</span>)
output = model(X)
<span class="keyword">print</span>(f<span class="string">"Output shape: {output.shape}"</span>) <span class="comment"># [1, 10]</span></div>
`
},
"activation": {
overview: `
<p>Activation functions are the <strong>secret sauce</strong> that gives neural networks their power. Without them, a neural network โ no matter how many layers โ would just be a fancy linear regression. They introduce <strong>non-linearity</strong>, which is what allows networks to learn complex patterns like recognizing faces, understanding language, or playing games.</p>
<div class="callout insight">
<div class="callout-title">๐ง The Big Idea</div>
Imagine trying to separate red and blue dots on a page using only straight lines. Without activation functions, that's all a neural network can do โ draw straight lines. With activation functions, the network can bend, curve, and twist those lines into any shape needed to perfectly separate the dots.
</div>
<h3>Why Do We Need Activation Functions?</h3>
<p>Consider stacking two linear layers:</p>
<div class="formula">
Layer 1: $$z_1 = W_1 x + b_1$$
Layer 2: $$z_2 = W_2 z_1 + b_2 = W_2(W_1 x + b_1) + b_2 = (W_2 W_1)x + (W_2 b_1 + b_2)$$
</div>
<p>The result is still just <strong>one linear transformation</strong>! Adding 100 more layers changes nothing โ it collapses into a single matrix multiplication. Activation functions break this linearity.</p>
<div class="info-box">
<div class="box-title">Learning Objectives</div>
<div class="box-content">
โ Understand <strong>why</strong> non-linearity is essential for deep learning<br>
โ Master 7 activation functions: Sigmoid, Tanh, ReLU, Leaky ReLU, ELU, Softmax, GELU<br>
โ Know the <strong>vanishing gradient problem</strong> and which activations solve it<br>
โ Learn when to use which activation (hidden layers vs. output layers)<br>
โ Implement all activations in PyTorch
</div>
</div>
<h3>Quick Comparison Table</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; text-align: left; color: var(--cyan);">Function</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Range</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Best For</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Issue</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);">
<td style="padding: 8px;">Sigmoid</td><td>(0, 1)</td><td>Binary output</td><td>Vanishing gradient</td>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);">
<td style="padding: 8px;">Tanh</td><td>(-1, 1)</td><td>Hidden layers (older models)</td><td>Vanishing gradient</td>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);">
<td style="padding: 8px;">ReLU</td><td>[0, โ)</td><td>Default for hidden layers</td><td>Dying neurons</td>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);">
<td style="padding: 8px;">Leaky ReLU</td><td>(-โ, โ)</td><td>When ReLU neurons die</td><td>Extra hyperparameter</td>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);">
<td style="padding: 8px;">GELU</td><td>โ(-0.17, โ)</td><td>Transformers (BERT, GPT)</td><td>Slightly slower</td>
</tr>
<tr>
<td style="padding: 8px;">Softmax</td><td>(0, 1) sums to 1</td><td>Multi-class output</td><td>Only for output layer</td>
</tr>
</table>
`,
concepts: `
<h3>1. Sigmoid โ The Classic Gate</h3>
<p>Sigmoid squashes any input into the range (0, 1). Think of it as a <strong>confidence score</strong> โ how sure is the network that something is true?</p>
<div class="formula" style="text-align:center;">$$\\sigma(z) = \\frac{1}{1 + e^{-z}}$$</div>
<div class="callout tip">
<div class="callout-title">๐ When to Use</div>
<strong>Only</strong> for the output layer of binary classification (yes/no, spam/not-spam). <strong>Never</strong> use in hidden layers โ it causes the vanishing gradient problem.
</div>
<p><strong>Why it vanishes:</strong> The maximum derivative of sigmoid is just 0.25 (at z=0). When you multiply gradients across 10 layers: 0.25ยนโฐ โ 0.00000095. The gradient essentially disappears, and the network stops learning.</p>
<h3>2. Tanh โ Centered Sigmoid</h3>
<p>Tanh is sigmoid's improvement โ it outputs values between (-1, 1), which means it's <strong>zero-centered</strong>. This helps with optimization because gradients don't all push in the same direction.</p>
<div class="formula" style="text-align:center;">$$\\tanh(z) = \\frac{e^z - e^{-z}}{e^z + e^{-z}} = 2\\sigma(2z) - 1$$</div>
<p>Fun fact: Tanh is just a rescaled sigmoid! It's better for hidden layers than sigmoid, but still vanishes for large |z|.</p>
<h3>3. ReLU โ The Game Changer</h3>
<p>ReLU (Rectified Linear Unit) is beautifully simple: output the input if positive, else output zero. This simplicity made deep learning practical.</p>
<div class="formula" style="text-align:center;">$$\\text{ReLU}(z) = \\max(0, z)$$</div>
<div class="callout insight">
<div class="callout-title">๐ฏ Why ReLU Changed Everything</div>
<strong>1. No vanishing gradient:</strong> Gradient is exactly 1 for positive inputs โ constant, never shrinking.<br>
<strong>2. Sparse activation:</strong> ~50% of neurons output 0. This sparsity makes networks efficient.<br>
<strong>3. Computationally cheap:</strong> Just a comparison โ no exponentials like sigmoid/tanh.<br>
<strong>4. Biologically plausible:</strong> Real neurons either fire or don't (all-or-nothing).
</div>
<p><strong>The Dying ReLU Problem:</strong> If a neuron's input is always negative (e.g., after a bad weight update), its gradient is permanently 0. That neuron is "dead" โ it will never activate again. Solution โ Leaky ReLU.</p>
<h3>4. Leaky ReLU & Parametric ReLU</h3>
<p>Instead of outputting 0 for negative inputs, Leaky ReLU allows a small slope (typically ฮฑ=0.01):</p>
<div class="formula" style="text-align:center;">$$\\text{LeakyReLU}(z) = \\begin{cases} z & z > 0 \\\\ \\alpha z & z \\leq 0 \\end{cases}$$</div>
<p><strong>Parametric ReLU (PReLU)</strong> makes ฮฑ a <em>learnable parameter</em> โ the network decides the optimal slope for negative values.</p>
<h3>5. GELU โ The Transformer's Choice</h3>
<p>GELU (Gaussian Error Linear Unit) is used in GPT, BERT, and almost all modern transformers. Unlike ReLU's hard cutoff at 0, GELU provides a <strong>smooth curve</strong>:</p>
<div class="formula" style="text-align:center;">$$\\text{GELU}(z) = z \\cdot \\Phi(z) \\approx 0.5z\\left(1 + \\tanh\\left[\\sqrt{\\frac{2}{\\pi}}(z + 0.044715z^3)\\right]\\right)$$</div>
<p>Where ฮฆ(z) is the cumulative distribution function of the standard normal. It "softly gates" the input based on how likely it is to be positive statistically.</p>
<h3>6. Swish / SiLU</h3>
<p>Discovered by Google through automated search. It's simply x times sigmoid(x):</p>
<div class="formula" style="text-align:center;">$$\\text{Swish}(z) = z \\cdot \\sigma(z) = \\frac{z}{1 + e^{-z}}$$</div>
<p>Swish is smooth, non-monotonic, and consistently outperforms ReLU in deep networks (>40 layers). Used in EfficientNet.</p>
<h3>7. Softmax โ Probability Distribution</h3>
<p>Softmax converts a vector of raw scores (logits) into a <strong>probability distribution</strong>. All outputs sum to 1.</p>
<div class="formula" style="text-align:center;">$$\\text{Softmax}(z_i) = \\frac{e^{z_i}}{\\sum_{j=1}^{K} e^{z_j}}$$</div>
<p><strong>Example:</strong> Logits [2.0, 1.0, 0.1] โ Softmax [0.659, 0.242, 0.099]. The network is 65.9% confident in class 0.</p>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Common Mistakes</div>
1. Using Softmax in hidden layers (it's for the final output layer only)<br>
2. Using Sigmoid for multi-class (use Softmax instead)<br>
3. Applying Softmax manually before CrossEntropyLoss in PyTorch (it already includes softmax!)
</div>
<h3>Decision Guide: Which Activation to Use?</h3>
<div class="info-box">
<div class="box-title">๐ฏ The Simple Rules</div>
<div class="box-content">
<strong>Hidden layers:</strong> Start with ReLU. If you see dying neurons, try Leaky ReLU. For Transformers, use GELU.<br><br>
<strong>Output layer โ Binary classification:</strong> Sigmoid (one neuron, output = probability)<br>
<strong>Output layer โ Multi-class:</strong> Softmax (K neurons, outputs = probabilities summing to 1)<br>
<strong>Output layer โ Regression:</strong> No activation (or Linear) โ let the output be any real number
</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ง Neural Network Design</div>
<div class="box-content">
Critical choice for every neural network - affects training speed, convergence, and final accuracy
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฏ Task-Specific Selection</div>
<div class="box-content">
Different tasks need different outputs: Sigmoid for binary, Softmax for multi-class, Linear for regression
</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ค Probable Interview Questions</div>
1. Why do we need activation functions?<br>
2. What is vanishing gradient?<br>
3. Why is ReLU preferred over sigmoid?<br>
4. What are dead neurons?<br>
5. Difference between ReLU and Leaky ReLU?<br>
6. Why softmax instead of sigmoid for multiclass?<br>
7. Why linear activation for regression output?<br>
8. Why GELU is used in transformers?<br>
9. Can activation function affect convergence speed?<br>
10. What happens if we remove activation functions?
</div>
`,
math: `
<h3>Derivatives: The Backprop Fuel</h3>
<p>Activation functions must be differentiable for backpropagation to work. Here are the key formulas rendered with proper math notation:</p>
<div class="formula" style="text-align:center;">
<strong>Sigmoid:</strong> $$\\sigma(z) = \\frac{1}{1 + e^{-z}}, \\quad \\sigma'(z) = \\sigma(z)(1 - \\sigma(z))$$
<strong>Tanh:</strong> $$\\tanh(z) = \\frac{e^z - e^{-z}}{e^z + e^{-z}}, \\quad \\tanh'(z) = 1 - \\tanh^2(z)$$
<strong>ReLU:</strong> $$\\text{ReLU}(z) = \\max(0, z), \\quad \\text{ReLU}'(z) = \\begin{cases} 1 & z > 0 \\ 0 & z \\leq 0 \\end{cases}$$
<strong>Leaky ReLU:</strong> $$\\text{LeakyReLU}(z) = \\begin{cases} z & z > 0 \\ \\alpha z & z \\leq 0 \\end{cases}, \\quad \\alpha = 0.01$$
<strong>Softmax:</strong> $$\\text{Softmax}(z_i) = \\frac{e^{z_i}}{\\sum_{j=1}^{K} e^{z_j}}$$
</div>
<div class="callout insight">
<div class="callout-title">๐ The Vanishing Gradient Problem (Proof)</div>
Each layer multiplies the gradient by $\\sigma'(z)$. For Sigmoid, $\\max(\\sigma') = 0.25$.<br>
For 10 Sigmoid layers: $$\\text{Total gradient} \\approx (0.25)^{10} \\approx 0.00000095$$
This is why <strong>ReLU</strong> ($\\text{gradient} = 1$) solved deep learning!
</div>
<h3>Python: Activation Functions</h3>
<span class="code-title">๐ activations.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn.functional <span class="keyword">as</span> F
x = torch.linspace(-<span class="number">5</span>, <span class="number">5</span>, <span class="number">100</span>)
<span class="comment"># All common activations</span>
relu = F.relu(x)
sigmoid = torch.sigmoid(x)
tanh = torch.tanh(x)
leaky = F.leaky_relu(x, negative_slope=<span class="number">0.01</span>)
gelu = F.gelu(x)
swish = x * torch.sigmoid(x) <span class="comment"># Swish = x * ฯ(x)</span>
<span class="comment"># Softmax (on logits)</span>
logits = torch.tensor([<span class="number">2.0</span>, <span class="number">1.0</span>, <span class="number">0.1</span>])
probs = F.softmax(logits, dim=<span class="number">0</span>)
<span class="keyword">print</span>(f<span class="string">"Softmax: {probs}"</span>) <span class="comment"># [0.659, 0.242, 0.099] โ sums to 1</span>
<span class="comment"># Custom activation in a model</span>
model = torch.nn.Sequential(
torch.nn.Linear(<span class="number">784</span>, <span class="number">256</span>),
torch.nn.GELU(), <span class="comment"># Used in Transformers</span>
torch.nn.Linear(<span class="number">256</span>, <span class="number">10</span>),
)</div>
`,
code: `
<h3>All Activation Functions in Python</h3>
<p>Let's implement every activation function from scratch using NumPy, then see the PyTorch equivalents.</p>
<span class="code-title">๐ activations_numpy.py</span><div class="code-block"><span class="keyword">import</span> numpy <span class="keyword">as</span> np
<span class="comment"># --- Activation Functions from Scratch ---</span>
<span class="keyword">def</span> <span class="function">sigmoid</span>(z):
<span class="string">"""Squashes input to (0, 1). Used for binary output."""</span>
<span class="keyword">return</span> <span class="number">1</span> / (<span class="number">1</span> + np.exp(-z))
<span class="keyword">def</span> <span class="function">sigmoid_derivative</span>(z):
s = sigmoid(z)
<span class="keyword">return</span> s * (<span class="number">1</span> - s) <span class="comment"># Max = 0.25 at z=0</span>
<span class="keyword">def</span> <span class="function">tanh_fn</span>(z):
<span class="string">"""Squashes input to (-1, 1). Zero-centered."""</span>
<span class="keyword">return</span> np.tanh(z)
<span class="keyword">def</span> <span class="function">relu</span>(z):
<span class="string">"""Max(0, z). Default for hidden layers."""</span>
<span class="keyword">return</span> np.maximum(<span class="number">0</span>, z)
<span class="keyword">def</span> <span class="function">leaky_relu</span>(z, alpha=<span class="number">0.01</span>):
<span class="string">"""Small gradient when z < 0. Fixes dying ReLU."""</span>
<span class="keyword">return</span> np.where(z > <span class="number">0</span>, z, alpha * z)
<span class="keyword">def</span> <span class="function">gelu</span>(z):
<span class="string">"""Used in BERT/GPT. Smooth approximation."""</span>
<span class="keyword">return</span> <span class="number">0.5</span> * z * (<span class="number">1</span> + np.tanh(np.sqrt(<span class="number">2</span>/np.pi) * (z + <span class="number">0.044715</span> * z**<span class="number">3</span>)))
<span class="keyword">def</span> <span class="function">swish</span>(z):
<span class="string">"""z * sigmoid(z). Self-gated. Used in EfficientNet."""</span>
<span class="keyword">return</span> z * sigmoid(z)
<span class="keyword">def</span> <span class="function">softmax</span>(z):
<span class="string">"""Converts logits to probabilities summing to 1."""</span>
exp_z = np.exp(z - np.max(z)) <span class="comment"># Numerical stability</span>
<span class="keyword">return</span> exp_z / exp_z.sum()
<span class="comment"># --- Demo ---</span>
z = np.linspace(-<span class="number">5</span>, <span class="number">5</span>, <span class="number">100</span>)
<span class="keyword">print</span>(<span class="string">"Sigmoid(0) ="</span>, sigmoid(<span class="number">0</span>)) <span class="comment"># 0.5</span>
<span class="keyword">print</span>(<span class="string">"ReLU(-3) ="</span>, relu(-<span class="number">3</span>)) <span class="comment"># 0</span>
<span class="keyword">print</span>(<span class="string">"Leaky ReLU(-3) ="</span>, leaky_relu(-<span class="number">3</span>)) <span class="comment"># -0.03</span>
<span class="keyword">print</span>(<span class="string">"Softmax([2,1,0.1]) ="</span>, softmax(np.array([<span class="number">2</span>, <span class="number">1</span>, <span class="number">0.1</span>])))
</div>
<h3>PyTorch Equivalents</h3>
<span class="code-title">๐ activations_pytorch.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> torch.nn.functional <span class="keyword">as</span> F
x = torch.randn(<span class="number">5</span>)
<span class="comment"># Functional API</span>
<span class="keyword">print</span>(<span class="string">"Sigmoid:"</span>, torch.sigmoid(x))
<span class="keyword">print</span>(<span class="string">"ReLU:"</span>, F.relu(x))
<span class="keyword">print</span>(<span class="string">"GELU:"</span>, F.gelu(x))
<span class="keyword">print</span>(<span class="string">"Softmax:"</span>, F.softmax(x, dim=<span class="number">0</span>))
<span class="comment"># In a model (Module API)</span>
model = nn.Sequential(
nn.Linear(<span class="number">784</span>, <span class="number">256</span>),
nn.ReLU(), <span class="comment"># Default for hidden layers</span>
nn.Linear(<span class="number">256</span>, <span class="number">128</span>),
nn.GELU(), <span class="comment"># Used in Transformers</span>
nn.Linear(<span class="number">128</span>, <span class="number">10</span>), <span class="comment"># No activation โ CrossEntropyLoss includes Softmax</span>
)
</div>
`
},
"conv-layer": {
overview: `
<p>Convolution is the <strong>fundamental operation</strong> that makes CNNs work. Instead of connecting every input pixel to every neuron (which would be billions of parameters for an image), a small filter (kernel) <strong>slides across the image</strong>, detecting local patterns like edges, corners, and textures.</p>
<div class="callout insight">
<div class="callout-title">๐ The Flashlight Analogy</div>
Imagine examining a large painting with a small flashlight in a dark room. You can only see a small patch at a time (the receptive field). As you move the flashlight across the painting (the sliding window), you notice patterns โ edges, curves, colors. A convolutional layer does exactly this, but with <strong>many flashlights simultaneously</strong>, each looking for a different pattern.
</div>
<h3>Key Concepts</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Kernel/Filter:</strong> A small matrix (typically 3ร3 or 5ร5) of learnable weights. Each filter detects one type of pattern.</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Stride:</strong> How many pixels the filter moves at each step. Stride=1 means move 1 pixel, stride=2 means skip every other position (halves output size).</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Padding:</strong> Adding zeros around the border so the output has the same spatial size as the input. 'same' padding with 3ร3 kernel means pad=1.</div></div>
<div class="list-item"><div class="list-num">4</div><div><strong>Feature Map:</strong> The output of applying one filter to the entire input. 32 filters = 32 feature maps = 32 channels in the output.</div></div>
<h3>Output Size Formula</h3>
<div class="formula" style="text-align:center;">
$$W_{out} = \\left\\lfloor \\frac{W_{in} + 2p - k}{s} \\right\\rfloor + 1$$
Where: $W_{in}$ = input size, $p$ = padding, $k$ = kernel size, $s$ = stride
</div>
<p><strong>Example:</strong> Input 32ร32, kernel 3ร3, stride 1, padding 1: Output = (32 + 2 - 3)/1 + 1 = <strong>32ร32</strong> (same size!)</p>
<h3>Parameter Efficiency: Why CNNs Beat Dense Layers</h3>
<div class="info-box">
<div class="box-title">๐ Parameter Comparison (32ร32 RGB Image โ 64 outputs)</div>
<div class="box-content">
<strong>Dense Layer:</strong> 32 ร 32 ร 3 ร 64 = <span style="color:#ff6b35">196,608 parameters</span><br>
<strong>Conv Layer (3ร3):</strong> 3 ร 3 ร 3 ร 64 + 64 = <span style="color:#00ff88">1,792 parameters</span><br>
<strong>Savings: 110x fewer parameters!</strong> This is thanks to weight sharing and local connectivity.
</div>
</div>
`,
concepts: `
<h3>Key Hyperparameters</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Kernel/Filter Size:</strong> Typically 3ร3 or 5ร5. Smaller = more layers needed, larger = more parameters</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Stride:</strong> Step size when sliding filter. Stride=1 (preserves size), Stride=2 (downsamples by 2ร)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Padding:</strong> Add zeros around borders. 'SAME' keeps size, 'VALID' shrinks output</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Number of Filters:</strong> Each filter learns different features. More filters = more capacity but slower</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Dilation:</strong> Spacing between kernel elements. Increases receptive field without adding parameters</div>
</div>
<div class="formula">
Output Size Formula:<br>
W_out = floor((W_in + 2รpadding - kernel_size) / stride) + 1<br>
H_out = floor((H_in + 2รpadding - kernel_size) / stride) + 1
</div>
`,
math: `
<h3>The Mathematical Operation: Cross-Correlation</h3>
<p>In deep learning, what we call "convolution" is mathematically "cross-correlation". It is a local dot product of the kernel and image patch.</p>
<div class="formula">
S(i, j) = (I * K)(i, j) = ฮฃ_m ฮฃ_n I(i+m, j+n) K(m, n)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Manual Convolution</div>
**Input (3x3):**<br>
[1 2 0]<br>
[0 1 1]<br>
[1 0 2]<br>
<br>
**Kernel (2x2):**<br>
[1 0]<br>
[0 1]<br>
<br>
**Calculation:**<br>
Step 1 (Top-Left): (1x1) + (2x0) + (0x0) + (1x1) = <strong>2</strong><br>
Step 2 (Top-Right): (2x1) + (0x0) + (1x0) + (1x1) = <strong>3</strong><br>
... Output is a 2x2 matrix.
</div>
<h3>Backprop through Conv</h3>
<p>Calculated using the same formula but with the kernel flipped vertically and horizontally (true convolution)!</p>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Feature Extraction</div>
<div class="box-content">
Early layers learn edges (Gabor-like filters), middle layers learn textures, deep layers learn specific object parts (eyes, wheels).
</div>
</div>
<div class="info-box">
<div class="box-title">๐จ Image Processing</div>
<div class="box-content">
Blurring, sharpening, and edge detection in Photoshop/GIMP are all done with 2D convolutions using fixed kernels.
</div>
</div>
`,
code: `
<h3>Convolution Math</h3>
<div class="formula" style="text-align:center;">
<strong>Output Size:</strong> $$W_{out} = \\left\\lfloor \\frac{W_{in} + 2p - k}{s} \\right\\rfloor + 1$$
Where $p$ = padding, $k$ = kernel size, $s$ = stride.
</div>
<h3>Complete CNN Classifier (CIFAR-10)</h3>
<span class="code-title">๐ cnn_cifar10.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> torchvision
<span class="keyword">import</span> torchvision.transforms <span class="keyword">as</span> T
<span class="comment"># --- Data Loading ---</span>
transform = T.Compose([T.ToTensor(), T.Normalize((<span class="number">0.5</span>,<span class="number">0.5</span>,<span class="number">0.5</span>), (<span class="number">0.5</span>,<span class="number">0.5</span>,<span class="number">0.5</span>))])
trainset = torchvision.datasets.CIFAR10(root=<span class="string">"./data"</span>, train=<span class="keyword">True</span>, download=<span class="keyword">True</span>, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=<span class="number">64</span>, shuffle=<span class="keyword">True</span>)
<span class="comment"># --- CNN Model ---</span>
<span class="keyword">class</span> <span class="class-name">CNN</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self):
<span class="builtin">super</span>().__init__()
self.features = nn.Sequential(
nn.Conv2d(<span class="number">3</span>, <span class="number">32</span>, <span class="number">3</span>, padding=<span class="number">1</span>), <span class="comment"># 32x32 โ 32x32</span>
nn.BatchNorm2d(<span class="number">32</span>), nn.ReLU(),
nn.MaxPool2d(<span class="number">2</span>), <span class="comment"># โ 16x16</span>
nn.Conv2d(<span class="number">32</span>, <span class="number">64</span>, <span class="number">3</span>, padding=<span class="number">1</span>), <span class="comment"># โ 16x16</span>
nn.BatchNorm2d(<span class="number">64</span>), nn.ReLU(),
nn.MaxPool2d(<span class="number">2</span>), <span class="comment"># โ 8x8</span>
nn.Conv2d(<span class="number">64</span>, <span class="number">128</span>, <span class="number">3</span>, padding=<span class="number">1</span>), <span class="comment"># โ 8x8</span>
nn.BatchNorm2d(<span class="number">128</span>), nn.ReLU(),
nn.AdaptiveAvgPool2d(<span class="number">1</span>), <span class="comment"># โ 1x1</span>
)
self.classifier = nn.Linear(<span class="number">128</span>, <span class="number">10</span>)
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
x = self.features(x)
x = x.view(x.size(<span class="number">0</span>), -<span class="number">1</span>)
<span class="keyword">return</span> self.classifier(x)
model = CNN()
<span class="keyword">print</span>(f<span class="string">"Parameters: {sum(p.numel() for p in model.parameters()):,}"</span>)</div>
`
},
"yolo": {
overview: `
<h3>What is YOLO?</h3>
<p>YOLO (You Only Look Once) treats object detection as a single regression problem, going directly from image pixels to bounding box coordinates and class probabilities in one forward pass.</p>
<h3>Why YOLO Over R-CNN?</h3>
<ul>
<li><strong>Speed:</strong> 45+ FPS (real-time) vs R-CNN's ~0.05 FPS</li>
<li><strong>Global Context:</strong> Sees entire image during training (fewer background errors)</li>
<li><strong>One Network:</strong> Unlike R-CNN's multi-stage pipeline</li>
<li><strong>End-to-End Training:</strong> Optimize detection directly</li>
</ul>
<div class="callout tip">
<div class="callout-title">โ
Advantages</div>
โข <strong>Lightning Fast:</strong> Real-time inference (YOLOv8 at 100+ FPS)<br>
โข <strong>Simple Architecture:</strong> Single network, easy to train<br>
โข <strong>Generalizes Well:</strong> Works on natural images and artwork<br>
โข <strong>Small Model Size:</strong> Can run on edge devices (mobile, IoT)
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Disadvantages</div>
โข <strong>Struggles with Small Objects:</strong> Grid limitation affects tiny items<br>
โข <strong>Localization Errors:</strong> Less precise than two-stage detectors<br>
โข <strong>Limited Objects per Cell:</strong> Can't detect many close objects<br>
โข <strong>Aspect Ratio Issues:</strong> Struggles with unusual object shapes
</div>
<h3>YOLO Evolution</h3>
<table>
<tr>
<th>Version</th>
<th>Year</th>
<th>Key Innovation</th>
<th>mAP</th>
</tr>
<tr>
<td>YOLOv1</td>
<td>2015</td>
<td>Original single-shot detector</td>
<td>63.4%</td>
</tr>
<tr>
<td>YOLOv3</td>
<td>2018</td>
<td>Multi-scale predictions</td>
<td>57.9% (faster)</td>
</tr>
<tr>
<td>YOLOv5</td>
<td>2020</td>
<td>PyTorch, Auto-augment</td>
<td>~50% (optimized)</td>
</tr>
<tr>
<td>YOLOv8</td>
<td>2023</td>
<td>Anchor-free, SOTA speed</td>
<td>53.9% (real-time)</td>
</tr>
</table>
`,
concepts: `
<h3>How YOLO Works (3 Steps)</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Grid Division:</strong> Divide image into SรS grid (e.g., 7ร7). Each cell predicts B bounding boxes</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Predictions Per Cell:</strong> Each box predicts (x, y, w, h, confidence) + class probabilities</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Non-Max Suppression:</strong> Remove duplicate detections, keep highest confidence boxes</div>
</div>
<div class="formula">
Output Tensor Shape (YOLOv1):<br>
S ร S ร (B ร 5 + C)<br>
Example: 7 ร 7 ร (2 ร 5 + 20) = 7 ร 7 ร 30<br>
<br>
Where:<br>
โข S = grid size (7)<br>
โข B = boxes per cell (2)<br>
โข 5 = (x, y, w, h, confidence)<br>
โข C = number of classes (20 for PASCAL VOC)
</div>
`,
applications: `
<h3>Industry Applications</h3>
<div class="info-box">
<div class="box-title">๐ Autonomous Vehicles</div>
<div class="box-content">
Real-time detection of pedestrians, vehicles, traffic signs, and lane markings for self-driving cars
</div>
</div>
<div class="info-box">
<div class="box-title">๐ญ Manufacturing</div>
<div class="box-content">
Quality control, defect detection on assembly lines, robot guidance, inventory management
</div>
</div>
<div class="info-box">
<div class="box-title">๐ก๏ธ Security & Surveillance</div>
<div class="box-content">
Intrusion detection, crowd monitoring, suspicious behavior analysis, license plate recognition
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฅ Medical Imaging</div>
<div class="box-content">
Tumor localization, cell counting, anatomical structure detection in X-rays/CT scans
</div>
</div>
`,
math: `
<h3>Intersection over Union (IoU)</h3>
<p>How do we measure if a predicted box is correct? We use the geometric ratio of intersection and union.</p>
<div class="formula">
IoU = Area of Overlap / Area of Union
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Manual IoU</div>
**Box A (GT):** [0,0,10,10] (Area=100)<br>
**Box B (Pred):** [5,5,15,15] (Area=100)<br>
1. **Intersection:** Area between [5,5] and [10,10] = 5x5 = 25<br>
2. **Union:** Area A + Area B - Intersection = 100 + 100 - 25 = 175<br>
3. **IoU:** 25 / 175 โ <strong>0.142</strong> (Poor match!)
</div>
<h3>YOLO Multi-Part Loss</h3>
<p>YOLO uses a composite loss function combining localization, confidence, and classification errors.</p>
<div class="formula">
L = ฮป_coord ฮฃ(Localization Loss) + ฮฃ(Confidence Loss) + ฮฃ(Classification Loss)
</div>
`
},
"transformers": {
overview: `
<p>The Transformer architecture, introduced in the 2017 paper <strong>"Attention Is All You Need"</strong>, is the foundation of modern AI. GPT, BERT, LLaMA, DALL-E, Stable Diffusion โ they all use Transformers. The key innovation: <strong>self-attention</strong>, which lets every token in a sequence directly attend to every other token.</p>
<div class="callout insight">
<div class="callout-title">๐ญ The Conference Room Analogy</div>
Imagine a meeting where everyone can hear everyone else simultaneously (unlike RNNs where information passes person-to-person in a chain). Each person (token) decides who to pay attention to based on relevance. The CEO might focus on the CFO for budget discussions but the CTO for technical decisions. This dynamic, context-dependent attention is what Transformers do.
</div>
<h3>Transformer Architecture</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Input Embedding + Positional Encoding:</strong> Convert tokens to vectors and add position information (since attention has no inherent sense of order)</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Multi-Head Self-Attention:</strong> Each head learns a different type of relationship. Head 1 might learn syntactic dependencies, Head 2 might learn semantic similarity.</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Feed-Forward Network:</strong> A two-layer MLP applied independently to each position. This is where most of the model's "knowledge" is stored.</div></div>
<div class="list-item"><div class="list-num">4</div><div><strong>Layer Normalization + Residual Connections:</strong> Applied around each sub-layer to stabilize training.</div></div>
<h3>Self-Attention: The Heart of Transformers</h3>
<div class="formula" style="text-align:center;">
$$\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$$
</div>
<p><strong>Q (Query):</strong> "What am I looking for?"<br>
<strong>K (Key):</strong> "What do I contain?"<br>
<strong>V (Value):</strong> "What information do I provide?"</p>
<p>The $\\sqrt{d_k}$ scaling prevents the dot products from becoming too large (which would make softmax saturate and kill gradients).</p>
<h3>Why Transformers Beat RNNs</h3>
<div class="info-box">
<div class="box-title">Speed & Power Comparison</div>
<div class="box-content">
<strong>Parallelization:</strong> RNNs process tokens sequentially (slow). Transformers process all tokens at once (fast). GPT-3 training would take 355 years with RNNs.<br>
<strong>Long-range dependencies:</strong> RNNs lose information over long sequences. Transformers can attend to any position directly.<br>
<strong>Scalability:</strong> Transformers scale to billions of parameters with predictable improvements (scaling laws).
</div>
</div>
`,
concepts: `
<h3>๐๏ธ Complete Architecture Overview</h3>
<p>The Transformer follows an <strong>Encoder-Decoder</strong> structure, but each component uses only attention mechanisms.</p>
<div class="info-box">
<div class="box-title">๐ฆ Model Hyperparameters (Base Model)</div>
<div class="box-content">
โข <strong>d_model = 512:</strong> Embedding dimension<br>
โข <strong>d_ff = 2048:</strong> Feed-forward hidden dimension<br>
โข <strong>h = 8:</strong> Number of attention heads<br>
โข <strong>d_k = d_v = 64:</strong> Key/Value dimensions (d_model / h)<br>
โข <strong>N = 6:</strong> Number of encoder AND decoder layers<br>
โข <strong>Total Parameters:</strong> ~65 million
</div>
</div>
<h3>Core Components</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Input Embedding:</strong> Convert tokens to d_model dimensional vectors, scaled by โd_model</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Positional Encoding:</strong> Add sinusoidal position signals so attention knows token order</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Multi-Head Self-Attention:</strong> h parallel attention heads, each computing different relationships</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Add & Norm:</strong> Residual connection followed by Layer Normalization</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Feed-Forward Network:</strong> Two linear layers with ReLU: FFN(x) = max(0, xWโ + bโ)Wโ + bโ</div>
</div>
<div class="list-item">
<div class="list-num">06</div>
<div><strong>Masked Attention (Decoder):</strong> Prevent attending to future tokens during training</div>
</div>
<div class="list-item">
<div class="list-num">07</div>
<div><strong>Encoder-Decoder Attention:</strong> Decoder attends to encoder output (cross-attention)</div>
</div>
<h3>Encoder Stack (N=6 layers)</h3>
<div class="formula">
Each encoder layer:<br>
sublayer_1 = LayerNorm(x + MultiHeadAttention(x, x, x))<br>
sublayer_2 = LayerNorm(sublayer_1 + FFN(sublayer_1))<br>
<br>
The encoder processes the input sequence bidirectionally.<br>
All positions can attend to all positions.
</div>
<h3>Decoder Stack (N=6 layers)</h3>
<div class="formula">
Each decoder layer has THREE sub-layers:<br>
1. Masked Self-Attention (prevent looking ahead)<br>
2. Encoder-Decoder Attention (attend to encoder output)<br>
3. Feed-Forward Network<br>
<br>
Output: LayerNorm(x + sublayer(x)) for each
</div>
<div class="callout insight">
<div class="callout-title">๐ก Key Insight: Residual Connections</div>
Every sub-layer has a residual connection: <strong>output = x + Sublayer(x)</strong><br>
This is crucial for:<br>
1. Training very deep models (gradient highway)<br>
2. Preserving information across layers<br>
3. Enabling optional "skipping" of transformations
</div>
`,
applications: `
<h3>๐ Revolutionary Applications</h3>
<div class="info-box">
<div class="box-title">๐ฌ Large Language Models</div>
<div class="box-content">
<strong>GPT-4, Claude, Gemini:</strong> The most capable AI systems ever built<br>
<strong>ChatGPT:</strong> 100M+ users in 2 months (fastest product adoption ever)<br>
<strong>BERT:</strong> Powers Google Search for 70+ languages
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Machine Translation (Original Use Case)</div>
<div class="box-content">
<strong>Google Translate:</strong> Switched to Transformers in 2017<br>
<strong>DeepL:</strong> Transformer-based, often beats Google<br>
<strong>NLLB-200:</strong> Meta's model translates 200 languages
</div>
</div>
<div class="info-box">
<div class="box-title">๐จ Multi-Modal AI</div>
<div class="box-content">
<strong>DALL-E 3, Midjourney, Stable Diffusion:</strong> Text-to-image generation<br>
<strong>GPT-4V, Gemini:</strong> Vision + Language understanding<br>
<strong>Whisper:</strong> State-of-the-art speech recognition<br>
<strong>Sora:</strong> Text-to-video generation
</div>
</div>
<div class="info-box">
<div class="box-title">๐งฌ Scientific Breakthroughs</div>
<div class="box-content">
<strong>AlphaFold 2:</strong> Solved 50-year protein folding problem (Nobel Prize 2024)<br>
<strong>ESMFold:</strong> Meta's protein predictor<br>
<strong>Drug Discovery:</strong> Accelerating molecule design
</div>
</div>
<div class="info-box">
<div class="box-title">๐ป Code & Development</div>
<div class="box-content">
<strong>GitHub Copilot:</strong> AI pair programmer (used by millions of devs)<br>
<strong>Claude Code, Cursor:</strong> AI coding assistants<br>
<strong>AlphaCode:</strong> Competitive programming solver
</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ Impact Statistics</div>
โข Paper has 100,000+ citations (one of most cited CS papers ever)<br>
โข Spawned multi-trillion dollar industry<br>
โข Every major AI lab now uses Transformer variants<br>
โข Considered the "ImageNet moment" for NLP
</div>
`,
math: `
<h3>๐ Paper & Pain: Complete Mathematical Derivation</h3>
<p>Let's derive every formula from the paper with step-by-step calculations.</p>
<h3>1. Scaled Dot-Product Attention</h3>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>Attention(Q, K, V) = softmax(QKแต / โdโ) V</strong>
</div>
<h4>Step-by-Step Derivation:</h4>
<div class="list-item">
<div class="list-num">01</div>
<div>
<strong>Create Q, K, V matrices:</strong><br>
Given input X โ โโฟหฃแต (n tokens, d dimensions)<br>
Q = XWแต , K = XWแดท, V = XWโฑฝ<br>
where Wแต , Wแดท โ โแตหฃแตแต and Wโฑฝ โ โแตหฃแตแต
</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div>
<strong>Compute Attention Scores:</strong><br>
scores = QKแต โ โโฟหฃโฟ<br>
Each score[i,j] = dot product of query_i and key_j<br>
Measures "how much should position i attend to position j"
</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div>
<strong>Scale by โdโ:</strong><br>
scaled_scores = QKแต / โdโ<br>
<em>Why scale?</em> Dot products grow with dimension. If dโ = 64:<br>
E[qแตk] = 0, Var[qแตk] = dโ = 64<br>
Without scaling, softmax saturates โ vanishing gradients!
</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div>
<strong>Apply Softmax:</strong><br>
attention_weights = softmax(scaled_scores) โ โโฟหฃโฟ<br>
Each row sums to 1 (probability distribution over positions)
</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div>
<strong>Weighted Sum of Values:</strong><br>
output = attention_weights ร V โ โโฟหฃแตแต<br>
Each output[i] is a weighted combination of all values
</div>
</div>
<h3>2. Multi-Head Attention</h3>
<div class="formula" style="background: rgba(255, 107, 53, 0.08); padding: 20px; border-radius: 8px;">
MultiHead(Q, K, V) = Concat(headโ, ..., headโ)Wแดผ<br>
where headแตข = Attention(QWแตขแต , KWแตขแดท, VWแตขโฑฝ)<br><br>
<strong>Paper's values:</strong><br>
h = 8 heads, dโ = dแตฅ = d_model/h = 512/8 = 64<br>
Wแดผ โ โโฝสฐยทแตแตโพหฃแตแตแตแตแตหก = โโตยนยฒหฃโตยนยฒ
</div>
<div class="callout insight">
<div class="callout-title">๐ก Why Multiple Heads?</div>
Each head can learn <strong>different relationships</strong>:<br>
โข Head 1: Syntactic dependencies (subject-verb)<br>
โข Head 2: Pronoun resolution (he โ John)<br>
โข Head 3: Semantic similarity<br>
โข Head 4-8: Other patterns<br>
<br>
<strong>Computation:</strong> Same cost as single-head with full dโ!<br>
8 heads ร 64 dim = 1 head ร 512 dim (same FLOPs)
</div>
<h3>3. Positional Encoding (Sinusoidal)</h3>
<div class="formula" style="background: rgba(46, 204, 113, 0.08); padding: 20px; border-radius: 8px;">
PE(pos, 2i) = sin(pos / 10000^(2i/d_model))<br>
PE(pos, 2i+1) = cos(pos / 10000^(2i/d_model))<br><br>
pos = position in sequence (0, 1, 2, ...)<br>
i = dimension index (0, 1, ..., d_model/2 - 1)
</div>
<h4>Worked Example (d_model = 4):</h4>
<div class="formula" style="background: var(--surface); padding: 15px;">
For position pos = 0:<br>
PE(0, 0) = sin(0) = 0<br>
PE(0, 1) = cos(0) = 1<br>
PE(0, 2) = sin(0) = 0<br>
PE(0, 3) = cos(0) = 1<br>
โ PEโ = [0, 1, 0, 1]<br><br>
For position pos = 1:<br>
PE(1, 0) = sin(1/10000โฐ) = sin(1) โ 0.84<br>
PE(1, 1) = cos(1/10000โฐ) = cos(1) โ 0.54<br>
PE(1, 2) = sin(1/100) โ 0.01<br>
PE(1, 3) = cos(1/100) โ 1.00<br>
โ PEโ = [0.84, 0.54, 0.01, 1.00]
</div>
<div class="callout warning">
<div class="callout-title">๐ Key Property: Relative Positions</div>
PE(pos + k) can be expressed as a <strong>linear function</strong> of PE(pos)!<br>
This allows the model to easily learn relative position attention.<br>
sin(a+b) = sin(a)cos(b) + cos(a)sin(b) โ
</div>
<h3>4. Feed-Forward Network</h3>
<div class="formula">
FFN(x) = max(0, xWโ + bโ)Wโ + bโ<br><br>
Wโ โ โแตแตแตแตแตหก หฃ แตแถ แถ = โโตยนยฒ หฃ ยฒโฐโดโธ (expand 4ร)<br>
Wโ โ โแตแถ แถ หฃ แตแตแตแตแตหก = โยฒโฐโดโธ หฃ โตยนยฒ (project back)<br><br>
This is a 2-layer MLP with ReLU, applied to each position independently.
</div>
<h3>5. Layer Normalization</h3>
<div class="formula">
LayerNorm(x) = ฮณ โ (x - ฮผ) / โ(ฯยฒ + ฮต) + ฮฒ<br><br>
ฮผ = (1/d) ฮฃแตข xแตข (mean across features)<br>
ฯยฒ = (1/d) ฮฃแตข (xแตข - ฮผ)ยฒ (variance)<br>
ฮณ, ฮฒ: learnable scale and shift (per feature)
</div>
<h3>6. Training Details from Paper</h3>
<div class="info-box">
<div class="box-title">โ๏ธ Optimization Settings</div>
<div class="box-content">
<strong>Optimizer:</strong> Adam with ฮฒโ=0.9, ฮฒโ=0.98, ฮต=10โปโน<br>
<strong>Learning Rate Schedule:</strong><br>
lr = d_modelโปโฐยทโต ร min(stepโปโฐยทโต, step ร warmup_stepsโปยนยทโต)<br>
warmup_steps = 4000<br><br>
<strong>Regularization:</strong><br>
โข Dropout = 0.1 (on sublayers and embeddings)<br>
โข Label Smoothing = 0.1
</div>
</div>
<h3>7. Complexity Analysis</h3>
<table>
<tr>
<th>Layer Type</th>
<th>Complexity per Layer</th>
<th>Sequential Ops</th>
<th>Max Path Length</th>
</tr>
<tr>
<td>Self-Attention</td>
<td>O(nยฒ ยท d)</td>
<td>O(1)</td>
<td>O(1)</td>
</tr>
<tr>
<td>Recurrent (RNN)</td>
<td>O(n ยท dยฒ)</td>
<td>O(n)</td>
<td>O(n)</td>
</tr>
<tr>
<td>Convolutional</td>
<td>O(k ยท n ยท dยฒ)</td>
<td>O(1)</td>
<td>O(logโ(n))</td>
</tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ฏ Key Insight from the Paper</div>
Self-attention is the only architecture with <strong>O(1) path length</strong><br>
between any two positions! This is why it handles long-range<br>
dependencies so well. The tradeoff is O(nยฒ) memory.
</div>
`,
visualization: `
<h3>๐จ Interactive Transformer Visualization</h3>
<p>Explore how attention mechanisms work in the Transformer architecture.</p>
<div class="viz-container">
<canvas id="transformerViz" width="800" height="600"></canvas>
</div>
<div class="viz-controls">
<button class="btn-viz" onclick="visualizeAttention()">Show Attention Weights</button>
<button class="btn-viz" onclick="visualizeMultiHead()">Multi-Head Attention</button>
<button class="btn-viz" onclick="visualizePositional()">Positional Encoding</button>
<button class="btn-viz" onclick="visualizeArchitecture()">Full Architecture</button>
</div>
<h3>Code: Self-Attention in PyTorch</h3>
<div class="formula" style="font-family: 'Courier New', monospace; font-size: 0.9rem; background: #1e1e2e; padding: 20px; border-radius: 8px; overflow-x: auto;">
<pre style="margin: 0; color: #cdd6f4;">
<span style="color: #89b4fa;">import</span> torch
<span style="color: #89b4fa;">import</span> torch.nn <span style="color: #89b4fa;">as</span> nn
<span style="color: #89b4fa;">import</span> torch.nn.functional <span style="color: #89b4fa;">as</span> F
<span style="color: #89b4fa;">import</span> math
<span style="color: #89b4fa;">class</span> <span style="color: #f9e2af;">ScaledDotProductAttention</span>(nn.Module):
<span style="color: #a6e3a1;">\"\"\"Attention(Q, K, V) = softmax(QK^T / sqrt(d_k)) V\"\"\"</span>
<span style="color: #89b4fa;">def</span> <span style="color: #f9e2af;">forward</span>(self, Q, K, V, mask=<span style="color: #fab387;">None</span>):
d_k = Q.size(-1)
<span style="color: #6c7086;"># Step 1: Compute attention scores</span>
scores = torch.matmul(Q, K.transpose(-2, -1))
<span style="color: #6c7086;"># Step 2: Scale by sqrt(d_k)</span>
scores = scores / math.sqrt(d_k)
<span style="color: #6c7086;"># Step 3: Apply mask (for decoder)</span>
<span style="color: #89b4fa;">if</span> mask <span style="color: #89b4fa;">is not</span> <span style="color: #fab387;">None</span>:
scores = scores.masked_fill(mask == <span style="color: #fab387;">0</span>, <span style="color: #fab387;">-1e9</span>)
<span style="color: #6c7086;"># Step 4: Apply softmax</span>
attention_weights = F.softmax(scores, dim=-1)
<span style="color: #6c7086;"># Step 5: Weighted sum of values</span>
output = torch.matmul(attention_weights, V)
<span style="color: #89b4fa;">return</span> output, attention_weights
<span style="color: #89b4fa;">class</span> <span style="color: #f9e2af;">MultiHeadAttention</span>(nn.Module):
<span style="color: #a6e3a1;">\"\"\"Multi-Head Attention from 'Attention Is All You Need'\"\"\"</span>
<span style="color: #89b4fa;">def</span> <span style="color: #f9e2af;">__init__</span>(self, d_model=<span style="color: #fab387;">512</span>, n_heads=<span style="color: #fab387;">8</span>):
<span style="color: #89b4fa;">super</span>().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.d_k = d_model // n_heads <span style="color: #6c7086;"># 64</span>
<span style="color: #6c7086;"># Linear projections for Q, K, V</span>
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, d_model)
self.W_V = nn.Linear(d_model, d_model)
self.W_O = nn.Linear(d_model, d_model)
self.attention = ScaledDotProductAttention()
<span style="color: #89b4fa;">def</span> <span style="color: #f9e2af;">forward</span>(self, Q, K, V, mask=<span style="color: #fab387;">None</span>):
batch_size = Q.size(<span style="color: #fab387;">0</span>)
<span style="color: #6c7086;"># 1. Linear projection and split into h heads</span>
Q = self.W_Q(Q).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
K = self.W_K(K).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
V = self.W_V(V).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
<span style="color: #6c7086;"># 2. Apply attention to all heads in parallel</span>
x, attn = self.attention(Q, K, V, mask)
<span style="color: #6c7086;"># 3. Concat and final linear projection</span>
x = x.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
<span style="color: #89b4fa;">return</span> self.W_O(x), attn
<span style="color: #89b4fa;">class</span> <span style="color: #f9e2af;">PositionalEncoding</span>(nn.Module):
<span style="color: #a6e3a1;">\"\"\"Sinusoidal Positional Encoding from the paper\"\"\"</span>
<span style="color: #89b4fa;">def</span> <span style="color: #f9e2af;">__init__</span>(self, d_model=<span style="color: #fab387;">512</span>, max_len=<span style="color: #fab387;">5000</span>):
<span style="color: #89b4fa;">super</span>().__init__()
pe = torch.zeros(max_len, d_model)
position = torch.arange(<span style="color: #fab387;">0</span>, max_len, dtype=torch.float).unsqueeze(<span style="color: #fab387;">1</span>)
div_term = torch.exp(torch.arange(<span style="color: #fab387;">0</span>, d_model, <span style="color: #fab387;">2</span>).float() * (-math.log(<span style="color: #fab387;">10000.0</span>) / d_model))
pe[:, <span style="color: #fab387;">0</span>::<span style="color: #fab387;">2</span>] = torch.sin(position * div_term)
pe[:, <span style="color: #fab387;">1</span>::<span style="color: #fab387;">2</span>] = torch.cos(position * div_term)
pe = pe.unsqueeze(<span style="color: #fab387;">0</span>) <span style="color: #6c7086;"># [1, max_len, d_model]</span>
self.register_buffer(<span style="color: #a6e3a1;">'pe'</span>, pe)
<span style="color: #89b4fa;">def</span> <span style="color: #f9e2af;">forward</span>(self, x):
<span style="color: #89b4fa;">return</span> x + self.pe[:, :x.size(<span style="color: #fab387;">1</span>)]
<span style="color: #6c7086;"># Example usage:</span>
batch_size, seq_len, d_model = <span style="color: #fab387;">2</span>, <span style="color: #fab387;">10</span>, <span style="color: #fab387;">512</span>
x = torch.randn(batch_size, seq_len, d_model)
mha = MultiHeadAttention(d_model=<span style="color: #fab387;">512</span>, n_heads=<span style="color: #fab387;">8</span>)
pe = PositionalEncoding(d_model=<span style="color: #fab387;">512</span>)
x = pe(x) <span style="color: #6c7086;"># Add positional encoding</span>
output, attn_weights = mha(x, x, x) <span style="color: #6c7086;"># Self-attention</span>
<span style="color: #89b4fa;">print</span>(f<span style="color: #a6e3a1;">"Output shape: {output.shape}"</span>) <span style="color: #6c7086;"># [2, 10, 512]</span>
<span style="color: #89b4fa;">print</span>(f<span style="color: #a6e3a1;">"Attention shape: {attn_weights.shape}"</span>) <span style="color: #6c7086;"># [2, 8, 10, 10]</span>
</pre>
</div>
<div class="callout tip">
<div class="callout-title">๐ Run This Code!</div>
Copy the code above and run it in a Jupyter notebook or Google Colab.<br>
Experiment with different d_model, n_heads, and seq_len values!
</div>
`,
code: `
<h3>Self-Attention (Scaled Dot-Product)</h3>
<div class="formula" style="text-align:center;">
$$\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$$
Where $Q = XW^Q$, $K = XW^K$, $V = XW^V$
</div>
<div class="formula" style="text-align:center;">
<strong>Multi-Head:</strong> $$\\text{MultiHead}(Q,K,V) = \\text{Concat}(head_1, ..., head_h)W^O$$
<strong>Positional Encoding:</strong> $$PE_{(pos, 2i)} = \\sin(pos / 10000^{2i/d}), \\quad PE_{(pos, 2i+1)} = \\cos(pos / 10000^{2i/d})$$
</div>
<span class="code-title">๐ transformer_from_scratch.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> math
<span class="keyword">class</span> <span class="class-name">SelfAttention</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self, d_model, n_heads):
<span class="builtin">super</span>().__init__()
self.d_k = d_model // n_heads
self.n_heads = n_heads
self.W_q = nn.Linear(d_model, d_model)
self.W_k = nn.Linear(d_model, d_model)
self.W_v = nn.Linear(d_model, d_model)
self.W_o = nn.Linear(d_model, d_model)
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
B, T, C = x.shape
Q = self.W_q(x).view(B, T, self.n_heads, self.d_k).transpose(<span class="number">1</span>,<span class="number">2</span>)
K = self.W_k(x).view(B, T, self.n_heads, self.d_k).transpose(<span class="number">1</span>,<span class="number">2</span>)
V = self.W_v(x).view(B, T, self.n_heads, self.d_k).transpose(<span class="number">1</span>,<span class="number">2</span>)
<span class="comment"># Scaled Dot-Product Attention</span>
scores = (Q @ K.transpose(-<span class="number">2</span>,-<span class="number">1</span>)) / math.sqrt(self.d_k)
attn = torch.softmax(scores, dim=-<span class="number">1</span>)
out = (attn @ V).transpose(<span class="number">1</span>,<span class="number">2</span>).reshape(B, T, C)
<span class="keyword">return</span> self.W_o(out)
<span class="comment"># Test</span>
attn = SelfAttention(d_model=<span class="number">512</span>, n_heads=<span class="number">8</span>)
x = torch.randn(<span class="number">2</span>, <span class="number">10</span>, <span class="number">512</span>) <span class="comment"># (batch, seq_len, d_model)</span>
<span class="keyword">print</span>(attn(x).shape) <span class="comment"># (2, 10, 512)</span></div>
`
},
"perceptron": {
overview: `
<h3>What is a Perceptron?</h3>
<p>The perceptron is the simplest neural network, invented in 1958. It's a binary linear classifier that makes predictions based on weighted inputs.</p>
<div class="callout tip">
<div class="callout-title">โ
Advantages</div>
โข Simple and fast<br>
โข Guaranteed convergence for linearly separable data<br>
โข Interpretable weights
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Key Limitation</div>
<strong>Cannot solve XOR:</strong> Limited to linear decision boundaries only
</div>
`,
concepts: `
<h3>How Perceptron Works</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Weighted Sum:</strong> z = wโxโ + wโxโ + ... + b</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Step Function:</strong> Output = 1 if z โฅ 0, else 0</div>
</div>
<div class="formula">
Learning Rule: w_new = w_old + ฮฑ(y_true - y_pred)x
</div>
`,
math: `
<h3>Perceptron Learning Algorithm</h3>
<p>The perceptron update rule is the simplest form of gradient descent.</p>
<div class="formula">
For each misclassified sample (x, y):<br>
w โ w + ฮฑ ร y ร x<br>
b โ b + ฮฑ ร y
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Manual Training</div>
<strong>Data:</strong> xโ = [1, 1], yโ = 1 | xโ = [0, 0], yโ = 0<br>
<strong>Initial:</strong> w = [0, 0], b = 0, ฮฑ = 1<br>
<br>
<strong>Iteration 1 (xโ):</strong><br>
z = 0ร1 + 0ร1 + 0 = 0 โ ลท = 1 โ (correct!)<br>
<br>
<strong>Iteration 2 (xโ):</strong><br>
z = 0ร0 + 0ร0 + 0 = 0 โ ลท = 1 โ (wrong! y=0)<br>
Update: w = [0,0] + 1ร(0-1)ร[0,0] = [0,0], b = 0 + 1ร(0-1) = -1<br>
<br>
Now z(xโ) = 0 + 0 - 1 = -1 โ ลท = 0 โ
</div>
<h3>Convergence Theorem</h3>
<div class="formula">
If data is linearly separable with margin ฮณ and ||x|| โค R,<br>
perceptron converges in at most (R/ฮณ)ยฒ updates.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Educational</div>
<div class="box-content">
Historical importance - first trainable neural model. Perfect for teaching ML fundamentals
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฌ Simple Classification</div>
<div class="box-content">
Linearly separable problems: basic pattern recognition, simple binary decisions
</div>
</div>
`
},
"mlp": {
overview: `
<h3>Multi-Layer Perceptron (MLP)</h3>
<p>MLP adds hidden layers between input and output, enabling non-linear decision boundaries and solving the XOR problem that single perceptrons cannot.</p>
<h3>Why MLPs?</h3>
<ul>
<li><strong>Universal Approximation:</strong> Can approximate any continuous function</li>
<li><strong>Non-Linear Learning:</strong> Solves complex problems</li>
<li><strong>Feature Extraction:</strong> Hidden layers learn hierarchical features</li>
</ul>
<div class="callout insight">
<div class="callout-title">๐ก The XOR Breakthrough</div>
Single perceptron: Cannot solve XOR<br>
MLP with 1 hidden layer (2 neurons): Solves XOR!<br>
This proves the power of depth.
</div>
`,
concepts: `
<h3>Architecture Components</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Input Layer:</strong> Raw features (no computation)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Hidden Layers:</strong> Extract progressively abstract features</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Output Layer:</strong> Final predictions</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Tabular Data</div>
<div class="box-content">Credit scoring, fraud detection, customer churn, sales forecasting</div>
</div>
<div class="info-box">
<div class="box-title">๐ญ Manufacturing</div>
<div class="box-content">Quality control, predictive maintenance, demand forecasting</div>
</div>
`,
math: `
<h3>Neural Network Forward Pass (Matrix Form)</h3>
<p>Vectorization is key to modern deep learning. We process entire layers as matrix multiplications.</p>
<div class="formula">
Layer 1: zโฝยนโพ = Wโฝยนโพx + bโฝยนโพ | aโฝยนโพ = ฯ(zโฝยนโพ)<br>
Layer 2: zโฝยฒโพ = Wโฝยฒโพaโฝยนโพ + bโฝยฒโพ | aโฝยฒโพ = ฯ(zโฝยฒโพ)<br>
...<br>
Layer L: ลท = Softmax(Wโฝแดธโพaโฝแดธโปยนโพ + bโฝแดธโพ)
</div>
<h3>Paper & Pain: Dimensionality Audit</h3>
<p>Understanding tensor shapes is the #1 skill for debugging neural networks.</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Input x:</strong> [n_features, 1]</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Weights Wโฝยนโพ:</strong> [n_hidden, n_features]</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Bias bโฝยนโพ:</strong> [n_hidden, 1]</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Solving XOR</div>
Input: [0,1], Target: 1<br>
Layer 1 (2 neurons):<br>
zโ = 10xโ + 10xโ - 5 | aโ = ฯ(zโ)<br>
zโ = 10xโ + 10xโ - 15 | aโ = ฯ(zโ)<br>
Layer 2 (1 neuron):<br>
y = ฯ(20aโ - 20aโ - 10)<br>
<strong>Try it on paper!</strong> This specific configuration correctly outputs XOR values.
</div>
`
},
"weight-init": {
overview: `
<h3>Weight Initialization Strategies</h3>
<table>
<tr>
<th>Method</th>
<th>Best For</th>
<th>Formula</th>
</tr>
<tr>
<td>Xavier/Glorot</td>
<td>Sigmoid, Tanh</td>
<td>N(0, โ(2/(n_in+n_out)))</td>
</tr>
<tr>
<td>He/Kaiming</td>
<td>ReLU</td>
<td>N(0, โ(2/n_in))</td>
</tr>
</table>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Never Initialize to Zero!</div>
All neurons learn identical features (symmetry problem)
</div>
`,
concepts: `
<h3>Key Principles</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Variance Preservation:</strong> Keep activation variance similar across layers</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Symmetry Breaking:</strong> Different weights force different features</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ Critical for Deep Networks</div>
<div class="box-content">
Proper initialization is essential for training networks >10 layers. Wrong init = training failure
</div>
</div>
<div class="info-box">
<div class="box-title">โก Faster Convergence</div>
<div class="box-content">
Good initialization reduces training time by 2-10ร, especially with modern optimizers
</div>
</div>
`,
math: `
<h3>The Variance Preservation Principle</h3>
<p>To prevent gradients from vanishing or exploding, we want the variance of the activations to remain constant across layers.</p>
<div class="formula">
For a linear layer: y = ฮฃ wแตขxแตข<br>
Var(y) = Var(ฮฃ wแตขxแตข) = ฮฃ Var(wแตขxแตข)<br>
Assuming w and x are independent with mean 0:<br>
Var(wแตขxแตข) = E[wแตขยฒ]E[xแตขยฒ] - E[wแตข]ยฒE[xแตข]ยฒ = Var(wแตข)Var(xแตข)<br>
So, Var(y) = n_in ร Var(w) ร Var(x)
</div>
<h3>1. Xavier (Glorot) Initialization</h3>
<p>Goal: Var(y) = Var(x) and Var(grad_out) = Var(grad_in)</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Forward Pass:</strong> n_in ร Var(w) = 1 โ Var(w) = 1/n_in</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Backward Pass:</strong> n_out ร Var(w) = 1 โ Var(w) = 1/n_out</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Compromise:</strong> Var(w) = 2 / (n_in + n_out)</div>
</div>
<h3>2. He (Kaiming) Initialization</h3>
<p>For ReLU activation, half the neurons are inactive (output 0), which halves the variance. We must compensate.</p>
<div class="formula">
Var(ReLU(y)) = 1/2 ร Var(y)<br>
To keep Var(ReLU(y)) = Var(x):<br>
1/2 ร n_in ร Var(w) = 1<br>
<strong>Var(w) = 2 / n_in</strong>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain Calculation</div>
If n_in = 256 and you use ReLU:<br>
Weight Std Dev = โ(2/256) = โ(1/128) โ <strong>0.088</strong><br>
Initializing with std=1.0 or std=0.01 would cause immediate failure in a deep net!
</div>
`
},
"loss": {
overview: `
<p>The loss function is the <strong>judge</strong> of your neural network. It measures how wrong the network's predictions are compared to the true answers. The entire goal of training is to <strong>minimize this number</strong>.</p>
<div class="callout insight">
<div class="callout-title">๐ฏ Why Loss Functions Matter</div>
The choice of loss function determines <em>what</em> the network optimizes for. Use the wrong loss and you'll get a network that's great at the wrong thing. For example, using MSE for classification gives equal penalty to being "slightly wrong" and "completely wrong" โ which isn't what you want.
</div>
<h3>Matching Loss to Task</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; text-align: left; color: var(--cyan);">Task</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Loss Function</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">Output Activation</th>
<th style="padding: 10px; text-align: left; color: var(--cyan);">PyTorch</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Regression</td><td>MSE / MAE / Huber</td><td>None (Linear)</td><td>nn.MSELoss()</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Binary Classification</td><td>Binary Cross-Entropy</td><td>Sigmoid</td><td>nn.BCEWithLogitsLoss()</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Multi-class</td><td>Categorical Cross-Entropy</td><td>Softmax (built-in)</td><td>nn.CrossEntropyLoss()</td></tr>
<tr><td style="padding: 8px;">Object Detection</td><td>Focal Loss + IoU Loss</td><td>Varies</td><td>Custom</td></tr>
</table>
`,
concepts: `
<h3>Common Loss Functions</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>MSE:</strong> (1/n)ฮฃ(y - ลท)ยฒ - Penalizes large errors, sensitive to outliers</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>MAE:</strong> (1/n)ฮฃ|y - ลท| - Robust to outliers, constant gradient, slower convergence</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Huber Loss:</strong> MSE when |error| โค ฮด, MAE otherwise. Best of both โ smooth + robust to outliers</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>BCE (Binary Cross-Entropy):</strong> -[yยทlog(ลท) + (1-y)ยทlog(1-ลท)] - Used with Sigmoid</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>CCE (Categorical Cross-Entropy):</strong> -ฮฃ yยทlog(ลท) - Used with Softmax for multi-class</div>
</div>
<div class="list-item">
<div class="list-num">06</div>
<div><strong>Hinge Loss:</strong> max(0, 1 - yยทลท) where y โ {-1, +1} - Margin-based, SVM-style</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ Task-Dependent Selection</div>
<div class="box-content">
Every ML task needs appropriate loss: classification (cross-entropy), regression (MSE/MAE), ranking (triplet loss)
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Custom Losses</div>
<div class="box-content">
Business-specific objectives: Focal Loss (imbalanced data), Dice Loss (segmentation), Contrastive Loss (similarity learning)
</div>
</div>
<h3>Loss Function Comparison</h3>
<table>
<tr><th>Loss</th><th>Type</th><th>Outlier Sensitivity</th><th>Key Property</th></tr>
<tr><td>MSE</td><td>Regression</td><td>High</td><td>Penalizes large errors heavily</td></tr>
<tr><td>MAE</td><td>Regression</td><td>Low</td><td>Robust, constant gradient</td></tr>
<tr><td>Huber</td><td>Regression</td><td>Medium</td><td>Smooth + robust (MSE+MAE combo)</td></tr>
<tr><td>BCE</td><td>Binary Class.</td><td>High</td><td>Strong gradients for wrong predictions</td></tr>
<tr><td>CCE</td><td>Multi-class</td><td>High</td><td>Outputs probabilities via Softmax</td></tr>
<tr><td>Hinge</td><td>Binary Class.</td><td>Medium</td><td>Margin-based, less probabilistic</td></tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ค Probable Interview Questions</div>
1. Difference between MSE and MAE?<br>
2. Why Huber loss is preferred sometimes?<br>
3. Why BCE with sigmoid?<br>
4. Why softmax with CCE?<br>
5. Why can't we use MSE for classification?<br>
6. What is Hinge loss and where is it used?<br>
7. Difference between loss function and evaluation metric?<br>
8. How does loss choice affect gradients?<br>
9. What is Focal Loss and when to use it?<br>
10. Can we design custom loss functions?
</div>
`,
math: `
<h3>Binary Cross-Entropy (BCE) Derivation</h3>
<p>Why do we use logs? BCE is derived from Maximum Likelihood Estimation (MLE) assuming a Bernoulli distribution.</p>
<div class="formula">
L(ลท, y) = -(y log(ลท) + (1-y) log(1-ลท))
</div>
<h3>Huber Loss (Smooth MAE)</h3>
<p>Combines MSE for small errors and MAE for large errors using threshold ฮด:</p>
<div class="formula">
L = ยฝ(y - ลท)ยฒ when |y - ลท| โค ฮด<br>
L = ฮด|y - ลท| - ยฝฮดยฒ otherwise
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Huber Intuition</div>
<strong>Small error (|error| โค ฮด):</strong> Behaves like MSE โ smooth, differentiable<br>
<strong>Large error (|error| > ฮด):</strong> Behaves like MAE โ doesn't blow up for outliers<br><br>
Best of both worlds! Used when data contains mild outliers.
</div>
<h3>Hinge Loss (SVM-style)</h3>
<div class="formula">
L = (1/n) ฮฃ max(0, 1 - yยทลท) where y โ {-1, +1}
</div>
<p>Margin-based loss: only penalizes predictions within the margin boundary. Used in SVMs and some neural network classifiers.</p>
<h3>Paper & Pain: Why not MSE for Classification?</h3>
<p>If we use MSE for sigmoid output, the gradient is:</p>
<div class="formula">
โL/โw = (ลท - y) <strong>ฯ'(z)</strong> x
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ The Saturation Problem</div>
If the model is very wrong (e.g., target 1, output 0.001), ฯ'(z) is near 0. <br>
The gradient vanishes, and the model <strong>stops learning!</strong>.
</div>
<h3>The BCE Advantage</h3>
<p>When using BCE, the ฯ'(z) term cancels out! The gradient becomes:</p>
<div class="formula" style="font-size: 1.2rem; color: #00d4ff;">
โL/โw = (ลท - y) x
</div>
<div class="list-item">
<div class="list-num">๐ก</div>
<div>This is beautiful: the gradient depends <strong>only on the error</strong> (ลท-y), not on how saturated the neuron is. This enables much faster training.</div>
</div>
`,
code: `
<h3>Loss Functions (MathJax)</h3>
<div class="formula" style="text-align:center;">
<strong>MSE (Regression):</strong> $$L = \\frac{1}{n}\\sum_{i=1}^{n}(y_i - \\hat{y}_i)^2$$
<strong>Binary Cross-Entropy:</strong> $$L = -\\frac{1}{n}\\sum[y\\log(\\hat{y}) + (1-y)\\log(1-\\hat{y})]$$
<strong>Categorical Cross-Entropy:</strong> $$L = -\\sum_{c=1}^{C} y_c \\log(\\hat{y}_c)$$
</div>
<span class="code-title">๐ loss_functions.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="comment"># --- MSE Loss (Regression) ---</span>
mse = nn.MSELoss()
pred = torch.tensor([<span class="number">2.5</span>, <span class="number">0.0</span>, <span class="number">2.1</span>])
target = torch.tensor([<span class="number">3.0</span>, <span class="number">-0.5</span>, <span class="number">2.0</span>])
<span class="keyword">print</span>(f<span class="string">"MSE: {mse(pred, target).item():.4f}"</span>)
<span class="comment"># --- Binary Cross-Entropy (Binary Classification) ---</span>
bce = nn.BCEWithLogitsLoss() <span class="comment"># includes sigmoid</span>
logits = torch.tensor([<span class="number">1.5</span>, <span class="number">-0.5</span>, <span class="number">2.0</span>])
labels = torch.tensor([<span class="number">1.0</span>, <span class="number">0.0</span>, <span class="number">1.0</span>])
<span class="keyword">print</span>(f<span class="string">"BCE: {bce(logits, labels).item():.4f}"</span>)
<span class="comment"># --- Cross-Entropy (Multi-class) ---</span>
ce = nn.CrossEntropyLoss() <span class="comment"># includes softmax internally</span>
logits = torch.tensor([[<span class="number">2.0</span>, <span class="number">1.0</span>, <span class="number">0.1</span>]])
target = torch.tensor([<span class="number">0</span>]) <span class="comment"># class index</span>
<span class="keyword">print</span>(f<span class="string">"CE: {ce(logits, target).item():.4f}"</span>)</div>
`
},
"optimizers": {
overview: `
<p>Optimizers are the <strong>algorithms that actually update the weights</strong> during training. Think of training as hiking down a foggy mountain โ you can't see the bottom (global minimum), but you can feel the slope beneath your feet (gradient). The optimizer decides <strong>how big a step to take</strong> and <strong>in which direction</strong>.</p>
<div class="callout insight">
<div class="callout-title">๐๏ธ The Mountain Analogy</div>
<strong>SGD</strong> = Walking downhill with fixed-size steps. Simple but slow.<br>
<strong>Momentum</strong> = Rolling a ball downhill โ it builds up speed on consistent slopes.<br>
<strong>Adam</strong> = A smart hiker with GPS who adapts step size per-dimension and remembers past terrain.
</div>
<h3>The Optimization Problem</h3>
<p>Training finds weights that <strong>minimize the loss function</strong>. The loss surface is a high-dimensional landscape with valleys (good), plateaus (stuck), and saddle points (tricky). Different optimizers navigate this landscape differently.</p>
<div class="formula" style="text-align:center;">
$$\\theta^* = \\arg\\min_{\\theta} L(\\theta; X, y)$$
</div>
<h3>Learning Rate: The Most Important Hyperparameter</h3>
<div class="info-box">
<div class="box-title">โก Learning Rate Effects</div>
<div class="box-content">
<strong>Too high (0.1):</strong> Overshoots the minimum, loss oscillates wildly or diverges to infinity<br>
<strong>Too low (0.00001):</strong> Converges extremely slowly, might get stuck in local minima<br>
<strong>Just right (0.001):</strong> Smooth convergence to a good minimum<br><br>
<strong>Rule of thumb:</strong> Start with lr=0.001 for Adam, lr=0.01 for SGD. Use learning rate schedulers (cosine annealing, step decay) to reduce lr during training.
</div>
</div>
`,
concepts: `
<h3>1. SGD (Stochastic Gradient Descent)</h3>
<p>The simplest optimizer. Compute the gradient on a mini-batch and take a step proportional to the learning rate:</p>
<div class="formula" style="text-align:center;">$$\\theta_{t+1} = \\theta_t - \\eta \\nabla L(\\theta_t)$$</div>
<p><strong>Problem:</strong> SGD oscillates in ravines (narrow valleys) and is slow on plateaus. It treats all parameters equally regardless of how frequently they update.</p>
<h3>2. SGD with Momentum</h3>
<p>Adds a "velocity" term that remembers past gradients. Like a ball rolling downhill โ it accelerates when the gradient consistently points the same direction:</p>
<div class="formula" style="text-align:center;">
$$v_t = \\gamma v_{t-1} + \\eta \\nabla L(\\theta_t)$$
$$\\theta_{t+1} = \\theta_t - v_t$$
</div>
<p>Where $\\gamma$ is the momentum coefficient (typically 0.9). This dampens oscillations and speeds convergence 2-10x.</p>
<h3>3. RMSprop</h3>
<p>Adapts the learning rate <strong>per parameter</strong>. Parameters with large gradients get smaller updates, and vice versa:</p>
<div class="formula" style="text-align:center;">
$$s_t = \\beta s_{t-1} + (1-\\beta)g_t^2$$
$$\\theta_{t+1} = \\theta_t - \\frac{\\eta}{\\sqrt{s_t + \\epsilon}} g_t$$
</div>
<p>Invented by Geoff Hinton in a Coursera lecture slide (never formally published!).</p>
<h3>4. Adam (Adaptive Moment Estimation) โญ</h3>
<p><strong>The default optimizer for most deep learning.</strong> It combines the best of Momentum (first moment) and RMSprop (second moment):</p>
<div class="formula" style="text-align:center;">
$$m_t = \\beta_1 m_{t-1} + (1-\\beta_1)g_t \\quad \\text{(1st moment: mean of gradients)}$$
$$v_t = \\beta_2 v_{t-1} + (1-\\beta_2)g_t^2 \\quad \\text{(2nd moment: variance of gradients)}$$
$$\\hat{m}_t = \\frac{m_t}{1-\\beta_1^t}, \\quad \\hat{v}_t = \\frac{v_t}{1-\\beta_2^t} \\quad \\text{(bias correction)}$$
$$\\theta_{t+1} = \\theta_t - \\frac{\\eta \\hat{m}_t}{\\sqrt{\\hat{v}_t} + \\epsilon}$$
</div>
<p>Default hyperparameters: $\\beta_1=0.9$, $\\beta_2=0.999$, $\\epsilon=10^{-8}$, $\\eta=0.001$.</p>
<h3>5. AdamW (Adam with Decoupled Weight Decay)</h3>
<p>Standard Adam applies weight decay <em>inside</em> the adaptive learning rate, which doesn't regularize properly. AdamW <strong>decouples</strong> weight decay from the gradient update:</p>
<div class="formula" style="text-align:center;">$$\\theta_{t+1} = (1 - \\lambda)\\theta_t - \\frac{\\eta \\hat{m}_t}{\\sqrt{\\hat{v}_t} + \\epsilon}$$</div>
<p><strong>Used by:</strong> GPT, BERT, LLaMA, and virtually all modern large models.</p>
<div class="callout tip">
<div class="callout-title">๐ฏ Which Optimizer Should You Use?</div>
<strong>Default:</strong> Adam or AdamW with lr=0.001<br>
<strong>Computer Vision:</strong> SGD with Momentum (lr=0.01, momentum=0.9) often generalizes better<br>
<strong>Large Language Models:</strong> AdamW with cosine lr schedule and warmup<br>
<strong>Fine-tuning pretrained:</strong> AdamW with very low lr (2e-5 to 5e-5)
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Training Acceleration</div>
<div class="box-content">
Modern optimizers (Adam) reduce training time by 5-10ร compared to basic SGD
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฏ Architecture-Specific</div>
<div class="box-content">
CNNs: SGD+Momentum | Transformers: AdamW | RNNs: RMSprop | Default: Adam
</div>
</div>
<h3>Optimizer Comparison</h3>
<table>
<tr><th>Optimizer</th><th>Key Idea</th><th>Problem</th></tr>
<tr><td>SGD</td><td>Simple, fast</td><td>Noisy convergence</td></tr>
<tr><td>Momentum</td><td>Smooths updates</td><td>Needs tuning</td></tr>
<tr><td>AdaGrad</td><td>Adaptive LR</td><td>LR shrinks too much</td></tr>
<tr><td>RMSProp</td><td>Fixes AdaGrad</td><td>No momentum</td></tr>
<tr><td><strong>Adam</strong></td><td><strong>Best of all</strong></td><td>Slightly more computation</td></tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ค Probable Interview Questions</div>
1. Difference between optimizer and gradient descent?<br>
2. Why does SGD oscillate?<br>
3. Why does AdaGrad fail in deep networks?<br>
4. How does RMSProp fix AdaGrad?<br>
5. Why is bias correction needed in Adam?<br>
6. What happens if learning rate is too high?<br>
7. When would you prefer SGD over Adam?<br>
8. What is momentum intuitively?<br>
9. Why is Adam the default choice?<br>
10. Can Adam overfit?
</div>
`,
math: `
<h3>Gradient Descent: The Foundation</h3>
<p>All optimizers are variations of gradient descent. The goal: minimize the loss function L(w).</p>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>w = w - ฮฑ ร โL(w)</strong>
</div>
<h3>1. Stochastic Gradient Descent (SGD)</h3>
<div class="formula">
w_{t+1} = w_t - ฮฑ ร โL(w_t)<br><br>
Where ฮฑ = learning rate (typically 0.01 - 0.1)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: SGD Step</div>
<strong>Current w:</strong> 2.0, <strong>Gradient:</strong> โL = 0.5, <strong>ฮฑ:</strong> 0.1<br><br>
w_new = 2.0 - 0.1 ร 0.5 = 2.0 - 0.05 = <strong>1.95</strong><br>
The weight moved slightly toward lower loss!
</div>
<h3>2. SGD with Momentum</h3>
<p>Adds a "velocity" term to accelerate through flat regions and dampen oscillations.</p>
<div class="formula">
v_{t+1} = ฮฒ ร v_t + โL(w_t)<br>
w_{t+1} = w_t - ฮฑ ร v_{t+1}<br><br>
Where ฮฒ = momentum coefficient (typically 0.9)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Momentum Step</div>
<strong>v_t:</strong> 0.3, <strong>โL:</strong> 0.5, <strong>ฮฒ:</strong> 0.9, <strong>ฮฑ:</strong> 0.1<br><br>
v_new = 0.9 ร 0.3 + 0.5 = 0.27 + 0.5 = <strong>0.77</strong><br>
w_new = w - 0.1 ร 0.77 = <strong>larger step!</strong><br>
Momentum accumulates past gradients for faster convergence.
</div>
<h3>3. AdaGrad (Adaptive Gradient)</h3>
<p>Adapts learning rate per-parameter based on how frequently each parameter is updated.</p>
<div class="formula">
<strong>Accumulated Gradient:</strong><br>
G_t = G_{t-1} + (โL)ยฒ<br><br>
<strong>Update Rule:</strong><br>
w_{t+1} = w_t - ฮท / โ(G_t + ฮต) ร โL<br><br>
Where ฮต = 1e-8 (numerical stability)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: AdaGrad Intuition</div>
<strong>Frequent parameters</strong> โ G_t grows fast โ learning rate shrinks<br>
<strong>Rare parameters</strong> โ G_t stays small โ learning rate stays large<br><br>
<strong>Problem:</strong> G_t only accumulates (never forgets), so learning rate keeps shrinking and training may stop early!<br>
<strong>This is exactly why RMSprop was invented โ</strong>
</div>
<h3>4. RMSprop (Root Mean Square Propagation)</h3>
<p>Fixes AdaGrad's shrinking problem by using a <strong>decaying average</strong> of recent squared gradients instead of summing all.</p>
<div class="formula">
v_t = ฮฒ ร v_{t-1} + (1-ฮฒ) ร (โL)ยฒ<br>
w_{t+1} = w_t - ฮฑ ร โL / โ(v_t + ฮต)<br><br>
ฮฒ = 0.9, ฮต = 1e-8 (numerical stability)
</div>
<h3>5. Adam (Adaptive Moment Estimation)</h3>
<p>Combines momentum (from SGD) AND adaptive learning rates (from RMSprop). The most popular optimizer.</p>
<div class="formula" style="background: rgba(255, 107, 53, 0.08); padding: 20px; border-radius: 8px;">
<strong>Step 1 - First Moment (Momentum):</strong><br>
m_t = ฮฒโ ร m_{t-1} + (1-ฮฒโ) ร โL<br><br>
<strong>Step 2 - Second Moment (RMSprop):</strong><br>
v_t = ฮฒโ ร v_{t-1} + (1-ฮฒโ) ร (โL)ยฒ<br><br>
<strong>Step 3 - Bias Correction:</strong><br>
mฬ_t = m_t / (1 - ฮฒโแต)<br>
vฬ_t = v_t / (1 - ฮฒโแต)<br><br>
<strong>Step 4 - Update:</strong><br>
w_{t+1} = w_t - ฮฑ ร mฬ_t / (โvฬ_t + ฮต)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Adam Step-by-Step</div>
<strong>Hyperparameters:</strong> ฮฑ=0.001, ฮฒโ=0.9, ฮฒโ=0.999, ฮต=1e-8<br>
<strong>t=2:</strong> โL = 0.5, mโ = 0.05, vโ = 0.00025<br><br>
mโ = 0.9 ร 0.05 + 0.1 ร 0.5 = 0.045 + 0.05 = 0.095<br>
vโ = 0.999 ร 0.00025 + 0.001 ร 0.25 = 0.000499<br><br>
mฬโ = 0.095 / (1 - 0.81) = 0.095 / 0.19 = 0.50<br>
vฬโ = 0.000499 / (1 - 0.998) = 0.2495<br><br>
ฮw = 0.001 ร 0.50 / (โ0.2495 + 1e-8) โ <strong>0.001</strong>
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Why Bias Correction?</div>
mโ = 0, vโ = 0 initialization biases early estimates toward zero.<br>
Dividing by (1 - ฮฒแต) compensates for this, especially in early training steps.
</div>
`,
code: `
<h3>Implementing Optimizers in PyTorch</h3>
<p>PyTorch provides all common optimizers in <code>torch.optim</code>. Here's a complete example showing how to use them and how to implement SGD from scratch.</p>
<span class="code-title">๐ optimizers_comparison.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> torch.optim <span class="keyword">as</span> optim
<span class="comment"># --- Simple model for demonstration ---</span>
model = nn.Sequential(
nn.Linear(<span class="number">784</span>, <span class="number">256</span>),
nn.ReLU(),
nn.Linear(<span class="number">256</span>, <span class="number">10</span>),
)
<span class="comment"># ====== 1. SGD ======</span>
optimizer = optim.SGD(model.parameters(), lr=<span class="number">0.01</span>)
<span class="comment"># ====== 2. SGD with Momentum ======</span>
optimizer = optim.SGD(model.parameters(), lr=<span class="number">0.01</span>, momentum=<span class="number">0.9</span>)
<span class="comment"># ====== 3. Adam (Most Common) ======</span>
optimizer = optim.Adam(model.parameters(), lr=<span class="number">0.001</span>)
<span class="comment"># ====== 4. AdamW (For Transformers/LLMs) ======</span>
optimizer = optim.AdamW(model.parameters(), lr=<span class="number">0.001</span>, weight_decay=<span class="number">0.01</span>)
<span class="comment"># ====== Training Loop ======</span>
criterion = nn.CrossEntropyLoss()
<span class="keyword">for</span> epoch <span class="keyword">in</span> range(<span class="number">10</span>):
<span class="keyword">for</span> X_batch, y_batch <span class="keyword">in</span> train_loader:
<span class="comment"># 1. Forward pass</span>
predictions = model(X_batch)
loss = criterion(predictions, y_batch)
<span class="comment"># 2. Backward pass</span>
optimizer.zero_grad() <span class="comment"># Clear old gradients</span>
loss.backward() <span class="comment"># Compute gradients</span>
<span class="comment"># 3. Update weights</span>
optimizer.step() <span class="comment"># Apply optimizer rule</span>
<span class="keyword">print</span>(f<span class="string">"Epoch {epoch+1}, Loss: {loss.item():.4f}"</span>)
</div>
<h3>SGD from Scratch</h3>
<p>Understanding what happens inside <code>optimizer.step()</code>:</p>
<span class="code-title">๐ sgd_from_scratch.py</span><div class="code-block"><span class="keyword">import</span> numpy <span class="keyword">as</span> np
<span class="keyword">class</span> <span class="function">SGD</span>:
<span class="string">"""Stochastic Gradient Descent with Momentum."""</span>
<span class="keyword">def</span> <span class="function">__init__</span>(self, params, lr=<span class="number">0.01</span>, momentum=<span class="number">0.9</span>):
self.params = params
self.lr = lr
self.momentum = momentum
self.velocities = [np.zeros_like(p) <span class="keyword">for</span> p <span class="keyword">in</span> params]
<span class="keyword">def</span> <span class="function">step</span>(self, gradients):
<span class="keyword">for</span> i, (param, grad) <span class="keyword">in</span> enumerate(zip(self.params, gradients)):
<span class="comment"># v = momentum * v_prev + lr * gradient</span>
self.velocities[i] = self.momentum * self.velocities[i] + self.lr * grad
<span class="comment"># param = param - velocity</span>
param -= self.velocities[i]
<span class="comment"># Usage</span>
W = np.random.randn(<span class="number">3</span>, <span class="number">3</span>) * <span class="number">0.01</span>
b = np.zeros(<span class="number">3</span>)
optimizer = SGD([W, b], lr=<span class="number">0.01</span>, momentum=<span class="number">0.9</span>)
<span class="comment"># After computing gradients dW, db:</span>
<span class="comment"># optimizer.step([dW, db])</span>
</div>
<h3>Learning Rate Schedulers</h3>
<span class="code-title">๐ lr_schedulers.py</span><div class="code-block"><span class="keyword">from</span> torch.optim.lr_scheduler <span class="keyword">import</span> StepLR, CosineAnnealingLR
optimizer = optim.Adam(model.parameters(), lr=<span class="number">0.001</span>)
<span class="comment"># Reduce LR by 0.1 every 30 epochs</span>
scheduler = StepLR(optimizer, step_size=<span class="number">30</span>, gamma=<span class="number">0.1</span>)
<span class="comment"># Cosine annealing (smooth decay, used in modern training)</span>
scheduler = CosineAnnealingLR(optimizer, T_max=<span class="number">100</span>, eta_min=<span class="number">1e-6</span>)
<span class="keyword">for</span> epoch <span class="keyword">in</span> range(<span class="number">100</span>):
train_one_epoch()
scheduler.step() <span class="comment"># Update learning rate</span>
<span class="keyword">print</span>(f<span class="string">"LR: {scheduler.get_last_lr()[0]:.6f}"</span>)
</div>
`
},
"backprop": {
overview: `
<p>Backpropagation is the algorithm that makes neural networks <strong>learn</strong>. It computes how much each weight in the network contributed to the error, then adjusts each weight accordingly. It's conceptually simple: <strong>the chain rule of calculus applied systematically from output to input</strong>.</p>
<div class="callout insight">
<div class="callout-title">๐ก The Factory Analogy</div>
Imagine a factory with 5 stations producing widgets. A quality inspector at the end finds defects. Backpropagation is like tracing the defect backward through each station to find which worker caused the problem and by how much. Each worker (weight) then adjusts their technique proportionally to their contribution to the defect.
</div>
<h3>The Algorithm in 4 Steps</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Forward Pass:</strong> Compute outputs layer by layer: $z^l = W^l a^{l-1} + b^l$, then $a^l = \\sigma(z^l)$</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Compute Loss:</strong> Compare prediction to truth: $L = \\text{CrossEntropy}(\\hat{y}, y)$</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Backward Pass:</strong> Compute $\\frac{\\partial L}{\\partial W^l}$ for every layer using the chain rule</div></div>
<div class="list-item"><div class="list-num">4</div><div><strong>Update:</strong> $W^l \\leftarrow W^l - \\eta \\frac{\\partial L}{\\partial W^l}$</div></div>
<h3>Computational Graphs</h3>
<p>Modern frameworks (PyTorch, TensorFlow) build a <strong>computational graph</strong> during the forward pass. Each operation becomes a node. During backprop, they traverse this graph in reverse, applying the chain rule at each node. This is called <strong>automatic differentiation (autograd)</strong>.</p>
<div class="info-box">
<div class="box-title">๐ Key Insight</div>
<div class="box-content">
Backprop's time complexity is <strong>O(W)</strong> โ proportional to the number of weights. This is the same as a forward pass! Computing all gradients is no more expensive than computing the prediction itself.
</div>
</div>
`,
concepts: `
<div class="formula">
Chain Rule:<br>
โL/โw = โL/โy ร โy/โz ร โz/โw<br>
<br>
For layer l:<br>
ฮดหก = (W^(l+1))^T ฮด^(l+1) โ ฯ'(z^l)<br>
โL/โW^l = ฮด^l (a^(l-1))^T
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ง Universal Training Method</div>
<div class="box-content">
Every modern neural network uses backprop - from CNNs to Transformers to GANs
</div>
</div>
<div class="info-box">
<div class="box-title">๐ง Automatic Differentiation</div>
<div class="box-content">
PyTorch, TensorFlow implement automatic backprop - you define forward pass, framework does backward
</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ค Probable Interview Questions</div>
1. What is the role of bias in a perceptron?<br>
2. Why can't we use MSE for classification?<br>
3. Difference between loss function and evaluation metric?<br>
4. Why is mini-batch GD preferred?<br>
5. Does backpropagation update weights?<br>
6. Can gradient descent work without backpropagation?<br>
7. What happens if learning rate is too high?<br>
8. How many times does forward propagation occur per epoch?<br>
9. What happens if we remove bias?<br>
10. What is the chain rule and why is it essential for backprop?
</div>
`,
math: `
<h3>The 4 Fundamental Equations of Backprop</h3>
<p>Backpropagation is essentially the chain rule applied iteratively. We define the error signal ฮด = โL/โz.</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Error at Output Layer (L):</strong><br>
ฮดแดธ = โโL โ ฯ'(zแดธ)<br>
<span class="formula-caption">Example for MSE: (aแดธ - y) โ ฯ'(zแดธ)</span></div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Error at Layer l (Backwards):</strong><br>
ฮดหก = ((Wหกโบยน)แต ฮดหกโบยน) โ ฯ'(zหก)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Gradient w.r.t Bias:</strong><br>
โL / โbหก = ฮดหก</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Gradient w.r.t Weights:</strong><br>
โL / โWหก = ฮดหก (aหกโปยน)แต</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain Walkthrough</div>
Suppose single neuron: z = wx + b, Loss L = (ฯ(z) - y)ยฒ/2<br>
1. <strong>Forward:</strong> z=2, a=ฯ(2)โ0.88, y=1, L=0.007<br>
2. <strong>Backward:</strong><br>
โL/โa = (a-y) = -0.12<br>
โa/โz = ฯ(z)(1-ฯ(z)) = 0.88 * 0.12 = 0.1056<br>
ฮด = โL/โz = -0.12 * 0.1056 = -0.01267<br>
<strong>โL/โw = ฮด * x</strong> | <strong>โL/โb = ฮด</strong>
</div>
`,
code: `
<h3>Backpropagation from Scratch</h3>
<p>A complete 2-layer neural network trained with backpropagation using only NumPy. This is the best way to truly understand how gradients flow.</p>
<span class="code-title">๐ backprop_numpy.py</span><div class="code-block"><span class="keyword">import</span> numpy <span class="keyword">as</span> np
np.random.seed(<span class="number">42</span>)
<span class="comment"># --- Data: XOR problem ---</span>
X = np.array([[<span class="number">0</span>,<span class="number">0</span>], [<span class="number">0</span>,<span class="number">1</span>], [<span class="number">1</span>,<span class="number">0</span>], [<span class="number">1</span>,<span class="number">1</span>]])
y = np.array([[<span class="number">0</span>], [<span class="number">1</span>], [<span class="number">1</span>], [<span class="number">0</span>]])
<span class="comment"># --- Initialize weights ---</span>
W1 = np.random.randn(<span class="number">2</span>, <span class="number">4</span>) * <span class="number">0.5</span> <span class="comment"># Input(2) -> Hidden(4)</span>
b1 = np.zeros((<span class="number">1</span>, <span class="number">4</span>))
W2 = np.random.randn(<span class="number">4</span>, <span class="number">1</span>) * <span class="number">0.5</span> <span class="comment"># Hidden(4) -> Output(1)</span>
b2 = np.zeros((<span class="number">1</span>, <span class="number">1</span>))
lr = <span class="number">0.5</span>
<span class="keyword">def</span> <span class="function">sigmoid</span>(z):
<span class="keyword">return</span> <span class="number">1</span> / (<span class="number">1</span> + np.exp(-z))
<span class="keyword">for</span> epoch <span class="keyword">in</span> range(<span class="number">10000</span>):
<span class="comment"># ====== FORWARD PASS ======</span>
z1 = X @ W1 + b1 <span class="comment"># (4, 4)</span>
a1 = sigmoid(z1) <span class="comment"># Hidden activation</span>
z2 = a1 @ W2 + b2 <span class="comment"># (4, 1)</span>
a2 = sigmoid(z2) <span class="comment"># Output prediction</span>
<span class="comment"># ====== COMPUTE LOSS (MSE) ======</span>
loss = np.mean((y - a2) ** <span class="number">2</span>)
<span class="comment"># ====== BACKWARD PASS (Chain Rule) ======</span>
<span class="comment"># dL/da2 = -2(y - a2) / n</span>
dL_da2 = -<span class="number">2</span> * (y - a2) / len(y)
<span class="comment"># da2/dz2 = sigmoid'(z2) = a2(1 - a2)</span>
dz2 = dL_da2 * a2 * (<span class="number">1</span> - a2)
<span class="comment"># Gradients for W2, b2</span>
dW2 = a1.T @ dz2 <span class="comment"># (4, 1)</span>
db2 = dz2.sum(axis=<span class="number">0</span>, keepdims=<span class="keyword">True</span>)
<span class="comment"># Propagate to hidden layer</span>
dz1 = (dz2 @ W2.T) * a1 * (<span class="number">1</span> - a1)
dW1 = X.T @ dz1 <span class="comment"># (2, 4)</span>
db1 = dz1.sum(axis=<span class="number">0</span>, keepdims=<span class="keyword">True</span>)
<span class="comment"># ====== UPDATE WEIGHTS ======</span>
W2 -= lr * dW2
b2 -= lr * db2
W1 -= lr * dW1
b1 -= lr * db1
<span class="keyword">if</span> epoch % <span class="number">2000</span> == <span class="number">0</span>:
<span class="keyword">print</span>(f<span class="string">"Epoch {epoch}: Loss = {loss:.6f}"</span>)
<span class="comment"># Final predictions</span>
<span class="keyword">print</span>(<span class="string">"Predictions:"</span>, a2.round(<span class="number">2</span>).flatten())
<span class="comment"># Expected: [0, 1, 1, 0] for XOR</span>
</div>
<h3>PyTorch Autograd (Automatic Backprop)</h3>
<span class="code-title">๐ backprop_pytorch.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="comment"># PyTorch handles backprop automatically via autograd!</span>
x = torch.tensor([<span class="number">2.0</span>], requires_grad=<span class="keyword">True</span>)
y = x**<span class="number">2</span> + <span class="number">3</span>*x + <span class="number">1</span> <span class="comment"># f(x) = x^2 + 3x + 1</span>
y.backward() <span class="comment"># Compute df/dx automatically</span>
<span class="keyword">print</span>(x.grad) <span class="comment"># 7.0 (derivative: 2x + 3, at x=2: 7)</span>
</div>
`
},
"regularization": {
overview: `
<p>Regularization prevents a neural network from <strong>memorizing the training data</strong> instead of learning general patterns. Without it, the network performs brilliantly on training data but fails on new data โ this is <strong>overfitting</strong>.</p>
<div class="callout insight">
<div class="callout-title">๐ The Exam Analogy</div>
Imagine a student who memorizes all answers in a textbook word-for-word but doesn't understand the concepts. They'll ace the exact same questions but fail on any new ones. Regularization forces the network to "understand" rather than "memorize."
</div>
<h3>Signs of Overfitting</h3>
<div class="info-box">
<div class="box-title">๐จ Red Flags</div>
<div class="box-content">
โ Training accuracy: 99%, Validation accuracy: 65% (huge gap)<br>
โ Training loss keeps decreasing but validation loss starts increasing<br>
โ Model works perfectly on seen data but fails on new examples<br>
โ Small dataset with a very large model (too many parameters for too little data)
</div>
</div>
<h3>Regularization Techniques</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>L2 Regularization (Weight Decay):</strong> Adds $\\lambda \\sum w^2$ to the loss. Penalizes large weights, encouraging smaller, more distributed weights.</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>L1 Regularization:</strong> Adds $\\lambda \\sum |w|$ to the loss. Encourages sparsity โ some weights become exactly zero.</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Dropout:</strong> During training, randomly sets neurons to 0 with probability $p$. Forces the network to be redundant โ no single neuron can be relied upon.</div></div>
<div class="list-item"><div class="list-num">4</div><div><strong>Data Augmentation:</strong> Artificially increase training data by applying random transformations (flips, rotations, crops). The best regularizer is <em>more data</em>.</div></div>
<div class="list-item"><div class="list-num">5</div><div><strong>Early Stopping:</strong> Monitor validation loss during training. Stop when it starts increasing even though training loss continues to decrease.</div></div>
<div class="list-item"><div class="list-num">6</div><div><strong>Batch Normalization:</strong> While not designed as regularization, BN has a regularizing effect because mini-batch statistics add noise.</div></div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ Best Practices</div>
<div class="box-content">
โข Start with Dropout (0.5) for hidden layers<br>
โข Add L2 if still overfitting (ฮป=0.01, 0.001)<br>
โข Always use Early Stopping<br>
โข Data Augmentation for images
</div>
</div>
<h3>Dropout vs Batch Normalization</h3>
<table>
<tr><th>Feature</th><th>Dropout</th><th>Batch Normalization</th></tr>
<tr><td>Purpose</td><td>Regularization</td><td>Faster training + mild regularization</td></tr>
<tr><td>Mechanism</td><td>Randomly drops neurons</td><td>Normalizes layer inputs</td></tr>
<tr><td>Training vs Test</td><td>Different behavior</td><td>Different behavior</td></tr>
<tr><td>Combined?</td><td colspan="2">Yes, use BatchNorm <em>before</em> Dropout</td></tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ค Probable Interview Questions</div>
1. Why can't we initialize all weights to zero?<br>
2. Difference between Xavier and He initialization?<br>
3. What is the vanishing gradient problem?<br>
4. How does Dropout prevent overfitting?<br>
5. Can we use Dropout at test time?<br>
6. Why is He initialization used with ReLU?<br>
7. What happens if weights are too large initially?<br>
8. Does Batch Normalization eliminate the need for Dropout?<br>
9. L1 vs L2 regularization โ when to use each?<br>
10. What is the exploding gradient problem and how to fix it?
</div>
`,
math: `
<h3>L2 Regularization (Weight Decay)</h3>
<p>Add a penalty proportional to the squared magnitude of weights.</p>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>L_total = L_data + ฮป ร ฮฃ wยฒ</strong>
</div>
<h4>Gradient with L2:</h4>
<div class="formula">
โL_total/โw = โL_data/โw + 2ฮปw<br><br>
Update rule becomes:<br>
w = w - ฮฑ(โL + 2ฮปw) = w(1 - 2ฮฑฮป) - ฮฑโL
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: L2 Effect</div>
<strong>Without L2:</strong> w = 5.0, โL = 0.1, ฮฑ = 0.1<br>
w_new = 5.0 - 0.1 ร 0.1 = 4.99<br><br>
<strong>With L2 (ฮป=0.01):</strong><br>
w_new = 5.0 ร (1 - 2ร0.1ร0.01) - 0.1 ร 0.1<br>
w_new = 5.0 ร 0.998 - 0.01 = 4.99 - 0.01 = <strong>4.98</strong><br><br>
The weight shrinks faster! Large weights shrink most.
</div>
<h3>L1 Regularization (Lasso)</h3>
<p>Adds penalty proportional to absolute value of weights - encourages sparsity.</p>
<div class="formula">
L_total = L_data + ฮป ร ฮฃ |w|<br><br>
Gradient: โL/โw = โL_data + ฮป ร sign(w)
</div>
<div class="callout tip">
<div class="callout-title">๐ก L1 vs L2</div>
โข <strong>L1:</strong> Creates sparse weights (many zeros) โ Feature selection<br>
โข <strong>L2:</strong> Small but non-zero weights โ More stable<br>
โข <strong>Elastic Net:</strong> ฮปโ|w| + ฮปโwยฒ (both!)
</div>
<h3>Dropout Mathematics</h3>
<p>Randomly set neurons to zero with probability p during training.</p>
<div class="formula">
<strong>Training:</strong><br>
r ~ Bernoulli(1-p) [mask of 0s and 1s]<br>
รฃ = a โ r [element-wise multiply]<br><br>
<strong>Inference (scaling):</strong><br>
รฃ = a ร (1-p) [scale by keep probability]
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Dropout Example</div>
<strong>Layer output (4 neurons):</strong> a = [1.0, 2.0, 0.5, 1.5]<br>
<strong>Dropout p = 0.5:</strong> r = [1, 0, 1, 0] (random mask)<br><br>
<strong>Training output:</strong> รฃ = [1.0, 0, 0.5, 0]<br>
<strong>Inference output:</strong> รฃ = [0.5, 1.0, 0.25, 0.75]<br><br>
During inference, we scale by (1-p)=0.5 to maintain expected value!
</div>
<h3>Why Dropout Works</h3>
<div class="formula">
Dropout โ Training an ensemble of 2โฟ sub-networks<br>
(where n = number of neurons that can be dropped)<br><br>
Each forward pass is a different architecture!
</div>
<h3>Weight Initialization Mathematics</h3>
<h4>Xavier Initialization (for Sigmoid/Tanh)</h4>
<div class="formula">
w ~ N(0, ฯยฒ) where ฯยฒ = 2 / (n_in + n_out)<br><br>
Goal: Keep Var(output) โ Var(input) across layers
</div>
<h4>He Initialization (for ReLU)</h4>
<div class="formula">
w ~ N(0, ฯยฒ) where ฯยฒ = 2 / n_in<br><br>
ReLU zeros out ~50% of activations, so variance is halved โ multiply by 2 to compensate!
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Why Zero Init Fails</div>
If all weights = 0, every neuron computes the <strong>same output</strong>.<br>
All gradients are <strong>identical</strong> โ All weights update the same way.<br>
Result: All neurons stay identical forever! The network is as good as <strong>1 neuron</strong>.<br><br>
<strong>Random Init:</strong> w ~ N(0, 0.01) works for shallow networks but gradients shrink exponentially in deep ones.<br>
<strong>Xavier:</strong> Calibrates variance based on layer width โ stable gradients for Sigmoid/Tanh.<br>
<strong>He:</strong> Accounts for ReLU zeroing out negative half โ default for modern networks.
</div>
`,
code: `
<h3>Regularization Techniques in PyTorch</h3>
<p>Complete examples of the most important regularization methods.</p>
<span class="code-title">๐ regularization_pytorch.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> torchvision.transforms <span class="keyword">as</span> T
<span class="comment"># ====== 1. Dropout ======</span>
<span class="keyword">class</span> <span class="function">ModelWithDropout</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self):
super().__init__()
self.net = nn.Sequential(
nn.Linear(<span class="number">784</span>, <span class="number">256</span>),
nn.ReLU(),
nn.Dropout(p=<span class="number">0.5</span>), <span class="comment"># 50% of neurons randomly zeroed</span>
nn.Linear(<span class="number">256</span>, <span class="number">128</span>),
nn.ReLU(),
nn.Dropout(p=<span class="number">0.3</span>), <span class="comment"># 30% dropout</span>
nn.Linear(<span class="number">128</span>, <span class="number">10</span>),
)
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
<span class="keyword">return</span> self.net(x)
model = ModelWithDropout()
model.train() <span class="comment"># Dropout active</span>
model.eval() <span class="comment"># Dropout disabled (IMPORTANT for inference!)</span>
<span class="comment"># ====== 2. L2 Regularization (Weight Decay) ======</span>
optimizer = torch.optim.Adam(
model.parameters(),
lr=<span class="number">0.001</span>,
weight_decay=<span class="number">0.01</span>, <span class="comment"># L2 penalty lambda</span>
)
<span class="comment"># ====== 3. Data Augmentation ======</span>
train_transforms = T.Compose([
T.RandomHorizontalFlip(p=<span class="number">0.5</span>),
T.RandomRotation(degrees=<span class="number">15</span>),
T.RandomResizedCrop(size=<span class="number">224</span>, scale=(<span class="number">0.8</span>, <span class="number">1.0</span>)),
T.ColorJitter(brightness=<span class="number">0.2</span>, contrast=<span class="number">0.2</span>),
T.ToTensor(),
T.Normalize(mean=[<span class="number">0.485</span>, <span class="number">0.456</span>, <span class="number">0.406</span>],
std=[<span class="number">0.229</span>, <span class="number">0.224</span>, <span class="number">0.225</span>]),
])
<span class="comment"># ====== 4. Early Stopping ======</span>
best_val_loss = float(<span class="string">'inf'</span>)
patience = <span class="number">5</span>
counter = <span class="number">0</span>
<span class="keyword">for</span> epoch <span class="keyword">in</span> range(<span class="number">100</span>):
train_loss = train_one_epoch()
val_loss = validate()
<span class="keyword">if</span> val_loss < best_val_loss:
best_val_loss = val_loss
counter = <span class="number">0</span>
torch.save(model.state_dict(), <span class="string">'best_model.pt'</span>)
<span class="keyword">else</span>:
counter += <span class="number">1</span>
<span class="keyword">if</span> counter >= patience:
<span class="keyword">print</span>(f<span class="string">"Early stopping at epoch {epoch}"</span>)
<span class="keyword">break</span>
</div>
`
},
"batch-norm": {
overview: `
<p>Batch Normalization (BatchNorm) was one of the most impactful innovations in deep learning. It <strong>normalizes the inputs to each layer</strong>, which stabilizes and dramatically accelerates training. Before BatchNorm, training deep networks was extremely difficult.</p>
<div class="callout insight">
<div class="callout-title">๐ฏ The Core Problem BatchNorm Solves</div>
As data flows through layers, the distribution of activations keeps shifting โ this is called <strong>Internal Covariate Shift</strong>. Each layer has to constantly adapt to new input distributions, slowing down training. BatchNorm fixes this by normalizing each layer's input to have mean=0 and variance=1.
</div>
<h3>How It Works (Step by Step)</h3>
<div class="formula" style="text-align:center;">
Given a mini-batch $B = \\{x_1, ..., x_m\\}$:<br><br>
<strong>Step 1 โ Mean:</strong> $$\\mu_B = \\frac{1}{m}\\sum_{i=1}^{m} x_i$$
<strong>Step 2 โ Variance:</strong> $$\\sigma_B^2 = \\frac{1}{m}\\sum_{i=1}^{m} (x_i - \\mu_B)^2$$
<strong>Step 3 โ Normalize:</strong> $$\\hat{x}_i = \\frac{x_i - \\mu_B}{\\sqrt{\\sigma_B^2 + \\epsilon}}$$
<strong>Step 4 โ Scale and Shift:</strong> $$y_i = \\gamma \\hat{x}_i + \\beta$$
</div>
<p>$\\gamma$ and $\\beta$ are <strong>learnable parameters</strong>. They allow the network to undo the normalization if that's optimal โ giving the network the flexibility to decide the best distribution for each layer.</p>
<div class="info-box">
<div class="box-title">Benefits of BatchNorm</div>
<div class="box-content">
โ <strong>Train 5-10x faster</strong> โ allows much higher learning rates<br>
โ <strong>Reduces sensitivity</strong> to weight initialization<br>
โ <strong>Acts as regularization</strong> โ mini-batch noise has a dropout-like effect<br>
โ <strong>Enables deeper networks</strong> โ reduces vanishing/exploding gradients
</div>
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ BatchNorm Gotchas</div>
<strong>1. Train vs. Eval:</strong> During training, BN uses mini-batch statistics. During inference, it uses running averages computed during training. Always call <code>model.eval()</code> before inference!<br>
<strong>2. Small batch sizes:</strong> BN performs poorly with batch_size < 16 because statistics are noisy. Use LayerNorm instead (as in Transformers).<br>
<strong>3. Placement:</strong> Usually placed <em>after</em> the linear/conv layer and <em>before</em> the activation: Linear โ BN โ ReLU.
</div>
`,
math: `
<h3>The 4 Steps of Batch Normalization</h3>
<p>Calculated per mini-batch B = {xโ, ..., xโ}:</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Mini-Batch Mean:</strong> ฮผ_B = (1/m) ฮฃ xแตข</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Mini-Batch Variance:</strong> ฯยฒ_B = (1/m) ฮฃ (xแตข - ฮผ_B)ยฒ</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Normalize:</strong> xฬแตข = (xแตข - ฮผ_B) / โ(ฯยฒ_B + ฮต)</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Scale and Shift:</strong> yแตข = ฮณ xฬแตข + ฮฒ</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Why ฮณ and ฮฒ?</div>
If we only normalized to (0,1), we might restrict the representation power of the network. <br>
ฮณ and ฮฒ allow the network to <strong>undo</strong> the normalization if that's optimal: <br>
If ฮณ = โ(ฯยฒ) and ฮฒ = ฮผ, we get the original data back!
</div>
`
},
"cv-intro": {
overview: `
<h3>Why Computer Vision Needs Special Architectures</h3>
<p><strong>Problem:</strong> Images have huge dimensionality</p>
<ul>
<li>224ร224 RGB image = 150,528 input features</li>
<li>Fully connected layer with 1000 neurons = 150M parameters!</li>
<li>Result: Overfitting, slow training, memory issues</li>
</ul>
<h3>Solution: Convolutional Neural Networks</h3>
<ul>
<li><strong>Weight Sharing:</strong> Same filter applied everywhere (1000x fewer parameters)</li>
<li><strong>Local Connectivity:</strong> Neurons see small patches</li>
<li><strong>Translation Invariance:</strong> Detect cat anywhere in image</li>
</ul>
`,
concepts: `
<h3>Why CNNs Beat Fully Connected</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Parameter Efficiency:</strong> 1000ร fewer parameters through weight sharing</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Translation Equivariance:</strong> Same object โ same activation regardless of position</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ธ Real-World CV</div>
<div class="box-content">
Face ID, medical imaging (MRI/CT), autonomous drone navigation, manufacturing defect detection, and satellite imagery analysis
</div>
</div>
`,
math: `
<h3>The Parameter Explosion Problem</h3>
<p>Why do standard Neural Networks fail on images? Let's calculate the parameters for a small image.</p>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: MLP vs Images</div>
1. **Input:** 224 ร 224 pixels with 3 color channels (RGB)<br>
2. **Input Size:** 224 ร 224 ร 3 = <strong>150,528 features</strong><br>
3. **Hidden Layer:** Suppose we want just 1000 neurons.<br>
4. **Matrix size:** [1000, 150528]<br>
5. **Total Weights:** 1000 ร 150528 โ <strong>150 Million parameters</strong> for just ONE layer!
</div>
<h3>The CNN Solution: Weight Sharing</h3>
<p>Instead of every neuron looking at every pixel, we use <strong>translation invariance</strong>. If an edge detector works in the top-left, it should work in the bottom-right.</p>
<div class="formula">
Total Params = (Kernel_H ร Kernel_W ร Input_Channels) ร Num_Filters<br>
<br>
For a 3x3 filter: (3 ร 3 ร 3) ร 64 = <strong>1,728 parameters</strong><br>
Reduction: 150M / 1.7k โ <strong>86,000ร more efficient!</strong>
</div>
`
},
"pooling": {
overview: `
<h3>Pooling Layers</h3>
<p>Pooling reduces spatial dimensions while retaining important information.</p>
<table>
<tr>
<th>Type</th>
<th>Operation</th>
<th>Use Case</th>
</tr>
<tr>
<td>Max Pooling</td>
<td>Take maximum value</td>
<td><strong>Most common</strong> - preserves strong activations</td>
</tr>
<tr>
<td>Average Pooling</td>
<td>Take average</td>
<td>Smoother, less common (used in final layers)</td>
</tr>
<tr>
<td>Global Pooling</td>
<td>Pool entire feature map</td>
<td>Replace FC layers (reduces parameters)</td>
</tr>
</table>
<div class="callout tip">
<div class="callout-title">โ
Benefits</div>
โข Reduces spatial size (faster computation)<br>
โข Adds translation invariance<br>
โข Prevents overfitting<br>
โข Typical: 2ร2 window, stride 2 (halves dimensions)
</div>
`,
concepts: `
<h3>Pooling Mechanics</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Downsampling:</strong> Reduces HรW by pooling factor (typically 2ร)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>No Learnable Parameters:</strong> Fixed operation (max/average)</div>
</div>
<div class="formula">
Example: 4ร4 input โ 2ร2 max pooling โ 2ร2 output
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ Standard CNN Component</div>
<div class="box-content">
Used after conv layers in AlexNet, VGG, and most classic CNNs to progressively reduce spatial dimensions
</div>
</div>
`,
math: `
<h3>Max Pooling: Winning Signal Selection</h3>
<p>Pooling operations are non-parametric (no weights). They simply select or average values within a local window.</p>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: 2x2 Max Pooling</div>
**Input (4x4):**<br>
[1 3 | 2 1]<br>
[5 1 | 0 2]<br>
-----------<br>
[1 1 | 8 2]<br>
[0 2 | 4 1]<br>
<br>
**Output (2x2):**<br>
Step 1: max(1, 3, 5, 1) = <strong>5</strong><br>
Step 2: max(2, 1, 0, 2) = <strong>2</strong><br>
Step 3: max(1, 1, 0, 2) = <strong>2</strong><br>
Step 4: max(8, 2, 4, 1) = <strong>8</strong><br>
**Final:** [5 2] / [2 8]
</div>
<h3>Backprop through Pooling</h3>
<div class="list-item">
<div class="list-num">๐ก</div>
<div><strong>Max Pooling:</strong> Gradient is routed ONLY to the neuron that had the maximum value. All others get 0.</div>
</div>
<div class="list-item">
<div class="list-num">๐ก</div>
<div><strong>Average Pooling:</strong> Gradient is distributed evenly among all neurons in the window.</div>
</div>
`
},
"cnn-basics": {
overview: `
<h3>CNN Architecture Pattern</h3>
<div class="formula">
Input โ [Conv โ ReLU โ Pool] ร N โ Flatten โ FC โ Softmax
</div>
<h3>Typical Layering Strategy</h3>
<ul>
<li><strong>Early Layers:</strong> Detect low-level features (edges, textures) - small filters (3ร3)</li>
<li><strong>Middle Layers:</strong> Combine into patterns, parts - more filters, same size</li>
<li><strong>Deep Layers:</strong> High-level concepts (faces, objects) - many filters</li>
<li><strong>Final FC Layers:</strong> Classification based on learned features</li>
</ul>
<div class="callout insight">
<div class="callout-title">๐ก Filter Progression</div>
Layer 1: 32 filters (edges)<br>
Layer 2: 64 filters (textures)<br>
Layer 3: 128 filters (patterns)<br>
Layer 4: 256 filters (parts)<br>
Common pattern: double filters after each pooling
</div>
`,
concepts: `
<h3>Module Design Principles</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Spatial Reduction:</strong> Progressively downsample (224โ112โ56โ28...)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Channel Expansion:</strong> Increase filters as spatial dims decrease</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ All Modern Vision Models</div>
<div class="box-content">
This pattern forms the backbone of ResNet, MobileNet, EfficientNet - fundamental CNN design
</div>
</div>
`,
math: `
<h3>1. The Golden Formula for Output Size</h3>
<p>Given Input (W), Filter Size (F), Padding (P), and Stride (S):</p>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0;">
Output Size = โ(W - F + 2P) / Sโ + 1
</div>
<h3>2. Parameter Count Calculation</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Parameters PER Filter:</strong> (F ร F ร C_in) + 1 (bias)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Total Parameters:</strong> N_filters ร ((F ร F ร C_in) + 1)</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain Calculation</div>
<strong>Input:</strong> 224x224x3 | <strong>Layer:</strong> 64 filters of 3x3 | <strong>Stride:</strong> 1 | <strong>Padding:</strong> 1<br>
1. <strong>Output Size:</strong> (224 - 3 + 2(1))/1 + 1 = 224 (Same Padding)<br>
2. <strong>Params:</strong> 64 * (3 * 3 * 3 + 1) = 64 * 28 = <strong>1,792 parameters</strong><br>
3. <strong>FLOPs:</strong> 224 * 224 * 1792 โ <strong>90 Million operations</strong> per image!
</div>
`,
code: `
<h3>Building a CNN from Scratch in PyTorch</h3>
<p>A complete CNN for classifying CIFAR-10 images (32ร32 RGB) into 10 categories.</p>
<span class="code-title">๐ cnn_cifar10.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">import</span> torch.nn.functional <span class="keyword">as</span> F
<span class="keyword">class</span> <span class="function">SimpleCNN</span>(nn.Module):
<span class="string">"""
Architecture:
Input (3, 32, 32)
-> Conv(3->32, 3x3) -> BN -> ReLU -> MaxPool(2x2) -> (32, 16, 16)
-> Conv(32->64, 3x3) -> BN -> ReLU -> MaxPool(2x2) -> (64, 8, 8)
-> Conv(64->128, 3x3) -> BN -> ReLU -> MaxPool(2x2) -> (128, 4, 4)
-> Flatten -> FC(2048, 256) -> Dropout -> FC(256, 10)
"""</span>
<span class="keyword">def</span> <span class="function">__init__</span>(self):
super().__init__()
<span class="comment"># Convolutional layers</span>
self.conv1 = nn.Conv2d(<span class="number">3</span>, <span class="number">32</span>, kernel_size=<span class="number">3</span>, padding=<span class="number">1</span>)
self.bn1 = nn.BatchNorm2d(<span class="number">32</span>)
self.conv2 = nn.Conv2d(<span class="number">32</span>, <span class="number">64</span>, kernel_size=<span class="number">3</span>, padding=<span class="number">1</span>)
self.bn2 = nn.BatchNorm2d(<span class="number">64</span>)
self.conv3 = nn.Conv2d(<span class="number">64</span>, <span class="number">128</span>, kernel_size=<span class="number">3</span>, padding=<span class="number">1</span>)
self.bn3 = nn.BatchNorm2d(<span class="number">128</span>)
self.pool = nn.MaxPool2d(<span class="number">2</span>, <span class="number">2</span>) <span class="comment"># Halves spatial dimensions</span>
<span class="comment"># Fully connected layers</span>
self.fc1 = nn.Linear(<span class="number">128</span> * <span class="number">4</span> * <span class="number">4</span>, <span class="number">256</span>)
self.dropout = nn.Dropout(<span class="number">0.5</span>)
self.fc2 = nn.Linear(<span class="number">256</span>, <span class="number">10</span>)
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
x = self.pool(F.relu(self.bn1(self.conv1(x)))) <span class="comment"># (B, 32, 16, 16)</span>
x = self.pool(F.relu(self.bn2(self.conv2(x)))) <span class="comment"># (B, 64, 8, 8)</span>
x = self.pool(F.relu(self.bn3(self.conv3(x)))) <span class="comment"># (B, 128, 4, 4)</span>
x = x.view(x.size(<span class="number">0</span>), -<span class="number">1</span>) <span class="comment"># Flatten</span>
x = self.dropout(F.relu(self.fc1(x)))
x = self.fc2(x)
<span class="keyword">return</span> x
<span class="comment"># --- Training Setup ---</span>
model = SimpleCNN()
<span class="keyword">print</span>(f<span class="string">"Total parameters: {sum(p.numel() for p in model.parameters()):,}"</span>)
<span class="comment"># Total parameters: ~600K (tiny by today's standards)</span>
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=<span class="number">0.001</span>)
</div>
`
},
"viz-filters": {
overview: `
<h3>What CNNs Learn</h3>
<p>CNN filters automatically learn hierarchical visual features:</p>
<h3>Layer-by-Layer Visualization</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Layer 1:</strong> Edges and colors (horizontal, vertical, diagonal lines)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Layer 2:</strong> Textures and patterns (corners, curves, simple shapes)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Layer 3:</strong> Object parts (eyes, wheels, windows)</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Layer 4-5:</strong> Whole objects (faces, cars, animals)</div>
</div>
`,
concepts: `
<h3>Visualization Techniques</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Activation Maximization:</strong> Find input that maximizes filter response</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Grad-CAM:</strong> Highlight important regions for predictions</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Model Interpretability</div>
<div class="box-content">
Understanding what CNNs learn helps debug failures, build trust, and improve architecture design
</div>
</div>
<div class="info-box">
<div class="box-title">๐จ Art & Style Transfer</div>
<div class="box-content">
Filter visualizations inspired neural style transfer (VGG features)
</div>
</div>
`,
math: `
<h3>Activation Maximization</h3>
<p>Find the input x* that maximizes a specific neuron's activation.</p>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>x* = argmax_x [a_ij(x) - ฮป||x||ยฒ]</strong>
</div>
<h4>Gradient Ascent on Input:</h4>
<div class="formula">
x_{t+1} = x_t + ฮฑ ร โa_ij/โx<br><br>
Where:<br>
โข a_ij = activation of neuron (i,j) in layer l<br>
โข ฮฑ = step size<br>
โข ฮป||x||ยฒ = regularization to keep input natural-looking
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Generate What a Filter "Sees"</div>
<strong>Goal:</strong> Find image that maximally activates Conv1 filter #5<br><br>
1. Start with x = random noise (224ร224ร3)<br>
2. Forward pass โ get activation aโ
at filter 5<br>
3. Backward pass: โaโ
/โx (gradient of activation w.r.t. input)<br>
4. Update: x = x + 0.01 ร โaโ
/โx<br>
5. Repeat 100-500 times<br><br>
<strong>Result:</strong> Image showing what pattern the filter detects!
</div>
<h3>Grad-CAM (Gradient-weighted Class Activation Mapping)</h3>
<p>Highlight which regions of the image were important for a specific class prediction.</p>
<div class="formula" style="background: rgba(255, 107, 53, 0.08); padding: 20px; border-radius: 8px;">
<strong>Step 1 - Global Average Pool the Gradients:</strong><br>
ฮฑโ = (1/Z) ร ฮฃแตข ฮฃโฑผ (โyแถ/โAโโฑสฒ)<br><br>
<strong>Step 2 - Weighted Sum of Feature Maps:</strong><br>
L_Grad-CAM = ReLU(ฮฃโ ฮฑโ ร Aโ)<br><br>
Where:<br>
โข yแถ = score for class c (before softmax)<br>
โข Aโ = k-th feature map of last conv layer<br>
โข ฮฑโ = importance weight of feature map k
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Grad-CAM Calculation</div>
<strong>Last conv layer:</strong> 14ร14ร512 feature maps<br>
<strong>Predicted class:</strong> "Dog" (class 5)<br><br>
1. Get gradient โyโ
/โA for all 512 feature maps<br>
2. Average each gradient map: ฮฑโ = 0.8, ฮฑโ = 0.1, ฮฑโ = 0.5...<br>
3. Weighted sum: L = 0.8รAโ + 0.1รAโ + 0.5รAโ + ...<br>
4. Apply ReLU (keep positive contributions only)<br>
5. Upsample to input size (14ร14 โ 224ร224)<br><br>
<strong>Result:</strong> Heatmap showing dog's face/body highlighted!
</div>
<h3>Saliency Maps (Vanilla Gradient)</h3>
<div class="formula">
Saliency(x) = |โyแถ/โx|<br><br>
Take absolute value of gradient of class score w.r.t. input pixels.
</div>
`
},
"lenet": {
overview: `
<h3>LeNet-5 (1998) - The Pioneer</h3>
<p>First successful CNN for digit recognition (MNIST). Introduced the Conv โ Pool โ Conv โ Pool pattern still used today.</p>
<h3>Architecture</h3>
<div class="formula">
Input 32ร32 โ Conv(6 filters, 5ร5) โ AvgPool โ Conv(16 filters, 5ร5) โ AvgPool โ FC(120) โ FC(84)โ FC(10)
</div>
<div class="callout insight">
<div class="callout-title">๐ Historical Impact</div>
โข Used by US Postal Service for zip code recognition<br>
โข Proved CNNs work for real-world tasks<br>
โข Template for modern architectures
</div>
`,
concepts: `
<h3>Key Innovations</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Layered Architecture:</strong> Hierarchical feature extraction</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Shared Weights:</strong> Convolutional parameter sharing</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">โ๏ธ Handwriting Recognition</div>
<div class="box-content">
USPS mail sorting, check processing, form digitization
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Educational Foundation</div>
<div class="box-content">
Perfect starting point for learning CNNs - simple enough to understand, complex enough to be useful
</div>
</div>
`,
math: `
<h3>LeNet-5 Complete Dimension Walkthrough</h3>
<p>Follow the tensor shapes through each layer of LeNet-5.</p>
<div class="formula" style="font-size: 1.0rem; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 20px; border-radius: 8px;">
<strong>Output Size Formula:</strong><br>
W_out = (W_in - K + 2P) / S + 1<br>
Where K=kernel, P=padding, S=stride
</div>
<h4>Layer-by-Layer Calculation:</h4>
<div class="list-item">
<div class="list-num">01</div>
<div>
<strong>Input:</strong> 32 ร 32 ร 1 (grayscale MNIST)<br>
<span class="formula-caption">Original: 28ร28, padded to 32ร32</span>
</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div>
<strong>C1 - Conv:</strong> 6 filters, 5ร5, stride 1, no padding<br>
Output: (32 - 5 + 0)/1 + 1 = <strong>28 ร 28 ร 6</strong><br>
Params: (5ร5ร1 + 1) ร 6 = <strong>156</strong>
</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div>
<strong>S2 - AvgPool:</strong> 2ร2, stride 2<br>
Output: 28/2 = <strong>14 ร 14 ร 6</strong><br>
Params: 0 (or 12 with learnable coefficients)
</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div>
<strong>C3 - Conv:</strong> 16 filters, 5ร5<br>
Output: (14 - 5)/1 + 1 = <strong>10 ร 10 ร 16</strong><br>
Params: (5ร5ร6 + 1) ร 16 = <strong>2,416</strong>
</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div>
<strong>S4 - AvgPool:</strong> 2ร2, stride 2<br>
Output: 10/2 = <strong>5 ร 5 ร 16</strong><br>
Params: 0
</div>
</div>
<div class="list-item">
<div class="list-num">06</div>
<div>
<strong>C5 - Conv:</strong> 120 filters, 5ร5<br>
Output: (5 - 5)/1 + 1 = <strong>1 ร 1 ร 120</strong><br>
Params: (5ร5ร16 + 1) ร 120 = <strong>48,120</strong>
</div>
</div>
<div class="list-item">
<div class="list-num">07</div>
<div>
<strong>F6 - FC:</strong> 120 โ 84<br>
Params: 120 ร 84 + 84 = <strong>10,164</strong>
</div>
</div>
<div class="list-item">
<div class="list-num">08</div>
<div>
<strong>Output - FC:</strong> 84 โ 10<br>
Params: 84 ร 10 + 10 = <strong>850</strong>
</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Total Parameter Count</div>
<strong>Convolutional Layers:</strong><br>
C1: 156 + C3: 2,416 + C5: 48,120 = 50,692<br><br>
<strong>Fully Connected Layers:</strong><br>
F6: 10,164 + Output: 850 = 11,014<br><br>
<strong>Total:</strong> โ <strong>61,706 parameters</strong><br><br>
Compare to AlexNet's 60 million - LeNet is 1000ร smaller!
</div>
<h3>Receptive Field Calculation</h3>
<div class="formula">
After C1 (5ร5): Each neuron sees 5ร5 pixels<br>
After S2: Sees 6ร6 pixels (pooling expands RF)<br>
After C3: Sees 14ร14 pixels<br>
After S4: Sees 16ร16 pixels<br>
After C5: <strong>Sees entire 32ร32 input!</strong>
</div>
`
},
"alexnet": {
overview: `
<h3>AlexNet (2012) - The Deep Learning Revolution</h3>
<p>Won ImageNet 2012 by huge margin (15.3% vs 26.2% error), igniting the deep learning revolution.</p>
<h3>Key Innovations</h3>
<ul>
<li><strong>ReLU Activation:</strong> Faster training than sigmoid/tanh</li>
<li><strong>Dropout:</strong> Prevents overfitting (p=0.5)</li>
<li><strong>Data Augmentation:</strong> Random crops/flips</li>
<li><strong>GPU Training:</strong> Used 2 GTX580 GPUs</li>
<li><strong>Deep:</strong> 8 layers (5 conv + 3 FC), 60M parameters</li>
</ul>
<div class="callout tip">
<div class="callout-title">๐ก Why So Important?</div>
First to show that deeper networks + more data + GPU compute = breakthrough performance
</div>
`,
concepts: `
<h3>Technical Contributions</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>ReLU:</strong> Solved vanishing gradients, enabled deeper networks</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Dropout:</strong> First major regularization for deep nets</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฏ ImageNet Challenge</div>
<div class="box-content">
Shattered records on 1000-class classification, proving deep learning superiority
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Industry Catalyst</div>
<div class="box-content">
Sparked AI renaissance - Google, Facebook, Microsoft pivoted to deep learning after AlexNet
</div>
</div>
`,
math: `
<h3>Paper & Pain: Parameter Counting</h3>
<p>Understanding AlexNet's 60M parameters:</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Conv Layers:</strong> Only ~2.3 Million parameters. They do most of the work with small memory!</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>FC Layers:</strong> Over **58 Million parameters**. The first FC layer (FC6) takes 4096 * (6*6*256) โ 37M params!</div>
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ The Design Flaw</div>
FC layers are the memory bottleneck. Modern models (ResNet, Inception) replace these with Global Average Pooling to save 90% parameters.
</div>
`
},
"vgg": {
overview: `
<h3>VGGNet (2014) - The Power of Depth</h3>
<p>VGG showed that depth matters - 16-19 layers using only small 3ร3 filters.</p>
`,
concepts: `
<h3>Small Filters, Receptive Field</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Uniformity:</strong> Uses 3ร3 filters everywhere with stride 1, padding 1.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Pooling Pattern:</strong> 2ร2 max pooling after every 2-3 conv layers.</div>
</div>
`,
math: `
<h3>The 5ร5 vs 3ร3+3ร3 Equivalence</h3>
<p>Why stack 3x3 filters instead of one large filter?</p>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Paramount Efficiency</div>
1. **Receptive Field:** Two 3x3 layers cover 5x5 area. Three 3x3 layers cover 7x7 area.<br>
2. **Param Count (C filters):**<br>
โข One 7x7 layer: 7ยฒ ร Cยฒ = 49Cยฒ parameters.<br>
โข Three 3x3 layers: 3 ร (3ยฒ ร Cยฒ) = 27Cยฒ parameters.<br>
**Result:** 45% reduction in weights for the SAME "view" of the image!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ผ๏ธ Feature Backbone</div>
VGG is the preferred architectural backbone for Neural Style Transfer and early GANs due to its simple, clean feature extraction properties.
</div>
`
},
"resnet": {
overview: `
<p>ResNet (Residual Networks) solved the <strong>degradation problem</strong> โ deeper networks were performing WORSE than shallower ones, not because of overfitting, but because gradients couldn't flow through many layers. The solution? <strong>Skip connections</strong> that let gradients bypass layers entirely.</p>
<div class="callout insight">
<div class="callout-title">๐๏ธ The Highway Analogy</div>
Imagine a highway with an exit ramp. If the road ahead is blocked (gradient vanishes), you can take the exit ramp and bypass the blockage. Skip connections are these exit ramps โ they guarantee that information (and gradients) can always flow through the network, even if some layers become useless.
</div>
<h3>The Residual Block</h3>
<p>Instead of learning the full mapping $H(x)$, the block learns the <strong>residual</strong> $F(x) = H(x) - x$. The output is:</p>
<div class="formula" style="text-align:center;">
$$y = F(x, \\{W_i\\}) + x$$
</div>
<p><strong>Why this works:</strong> If the optimal mapping is close to identity (common in deep networks), it's much easier for the network to learn $F(x) \\approx 0$ than to learn $H(x) \\approx x$. Pushing weights toward zero is easy; learning a complex identity mapping is hard.</p>
<h3>ResNet Variants</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Model</th>
<th style="padding: 10px; color: var(--cyan);">Layers</th>
<th style="padding: 10px; color: var(--cyan);">Parameters</th>
<th style="padding: 10px; color: var(--cyan);">Top-1 Accuracy</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">ResNet-18</td><td>18</td><td>11.7M</td><td>69.8%</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">ResNet-34</td><td>34</td><td>21.8M</td><td>73.3%</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">ResNet-50</td><td>50</td><td>25.6M</td><td>76.1%</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">ResNet-101</td><td>101</td><td>44.5M</td><td>77.4%</td></tr>
<tr><td style="padding: 8px;">ResNet-152</td><td>152</td><td>60.2M</td><td>78.3%</td></tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ก Impact</div>
ResNet won the 2015 ImageNet competition with 152 layers and 3.6% top-5 error โ the first time a neural network surpassed human performance (5.1%). ResNet concepts are now used in virtually every modern architecture.
</div>
`,
concepts: `
<h3>Identity & Projection Shortcuts</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Identity Shortcut:</strong> Used when dimensions match. y = F(x, {W}) + x</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Projection Shortcut (1ร1 Conv):</strong> Used when dimensions change. y = F(x, {W}) + W_s x</div>
</div>
`,
math: `
<h3>The Vanishing Gradient Solution</h3>
<p>Why do skip connections help? Let's differentiate the output y = F(x) + x:</p>
<div class="formula">
โy/โx = โF/โx + 1
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Gradient Flow</div>
The "+1" term acts as a **gradient highway**. Even if the weights in F(x) are small (causing โF/โx โ 0), the gradient can still flow through the +1 term. <br>
This prevents the gradient from vanishing even in networks with 1000+ layers!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐๏ธ Modern Vision Backbones</div>
<div class="box-content">ResNet is the default starting point for nearly all computer vision tasks today (Mask R-CNN, YOLO, etc.).</div>
</div>
`,
code: `
<h3>Skip Connection Math</h3>
<div class="formula" style="text-align:center;">
$$\\mathcal{H}(x) = \\mathcal{F}(x) + x$$
The network learns the <strong>residual</strong> $\\mathcal{F}(x) = \\mathcal{H}(x) - x$ instead of the full mapping.
</div>
<span class="code-title">๐ resnet_pytorch.py</span><div class="code-block"><span class="keyword">import</span> torchvision.models <span class="keyword">as</span> models
<span class="comment"># Load pretrained ResNet-50</span>
model = models.resnet50(weights=<span class="string">"IMAGENET1K_V2"</span>)
<span class="keyword">print</span>(f<span class="string">"Parameters: {sum(p.numel() for p in model.parameters()):,}"</span>)
<span class="comment"># Output: 25,557,032 parameters</span></div>
`
},
"inception": {
overview: `
<h3>Inception/GoogLeNet (2014) - Going Wider</h3>
<p>Instead of going deeper, Inception modules go wider - using multiple filter sizes in parallel.</p>
<h3>Inception Module</h3>
<div class="formula">
Input โ [1ร1 conv] โ [3ร3 conv] โ [5ร5 conv] โ [3ร3 pool] โ Concatenate
</div>
`,
concepts: `
<h3>Core Innovations</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>1ร1 Bottlenecks:</strong> Dimensionality reduction before expensive convolutions.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Auxiliary Classifiers:</strong> Used during training to combat gradient vanishing in middle layers.</div>
</div>
`,
math: `
<h3>1ร1 Convolution Math (Network-in-Network)</h3>
<p>A 1ร1 convolution acts like a channel-wise MLP. It maps input channels C to output channels C' using 1ร1รC parameters per filter.</p>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Compression</div>
Input: 28x28x256 | Target: 28x28x512 with 3x3 Filters.<br>
**Direct:** 512 * (3*3*256) โ 1.1 Million params.<br>
**Inception (1x1 bottleneck to 64):**<br>
Step 1 (1x1): 64 * (1*1*256) = 16k params.<br>
Step 2 (3x3): 512 * (3*3*64) = 294k params.<br>
**Total:** 310k params. **~3.5ร reduction in parameters!**
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐๏ธ Computational Efficiency</div>
Inception designs are optimized for running deep networks on limited compute budgets.
</div>
`
},
"mobilenet": {
overview: `
<h3>MobileNet - CNNs for Mobile Devices</h3>
<p>Designed for mobile/embedded vision using depthwise separable convolutions.</p>
<h3>Depthwise Separable Convolution</h3>
<div class="formula">
Standard Conv = Depthwise Conv + Pointwise (1ร1) Conv
</div>
<h3>Computation Reduction</h3>
<table>
<tr>
<th>Method</th>
<th>Parameters</th>
<th>FLOPs</th>
</tr>
<tr>
<td>Standard 3ร3 Conv</td>
<td>3ร3รC_inรC_out</td>
<td>High</td>
</tr>
<tr>
<td>Depthwise Separable</td>
<td>3ร3รC_in + C_inรC_out</td>
<td><strong>8-9ร less!</strong></td>
</tr>
</table>
<div class="callout tip">
<div class="callout-title">โ
Applications</div>
โข Real-time mobile apps (camera filters, AR)<br>
โข Edge devices (drones, IoT)<br>
โข Latency-critical systems<br>
โข Good accuracy with 10-20ร speedup
</div>
`,
concepts: `
<h3>Efficiency Factors</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Width Multiplier (ฮฑ):</strong> Thins the network by reducing channels.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Resolution Multiplier (ฯ):</strong> Reduces input image size.</div>
</div>
`,
math: `
<h3>Depthwise Separable Math</h3>
<p>Standard convolution complexity: Fยฒ ร C_in ร C_out ร H ร W</p>
<p>Separable complexity: (Fยฒ ร C_in + C_in ร C_out) ร H ร W</p>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: The 9ร Speedup</div>
Reduction ratio is roughly: 1/C_out + 1/Fยฒ. <br>
For 3x3 filters (F=3): Reduction is roughly **1/9th** the computation of standard conv!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฑ Edge Devices</div>
<div class="box-content">Real-time object detection on smartphones, web browsers (TensorFlow.js), and IoT devices.</div>
</div>
`
},
"transfer-learning": {
overview: `
<p>Transfer learning is the most <strong>practical technique</strong> in modern deep learning. Instead of training from scratch (which requires millions of images and weeks of GPU time), you <strong>reuse a model already trained on a large dataset</strong> (like ImageNet's 14M images) and fine-tune it for your specific task.</p>
<div class="callout insight">
<div class="callout-title">๐ The Education Analogy</div>
Transfer learning is like a medical student who already learned anatomy, chemistry, and biology (pretrained knowledge). Instead of re-learning everything from scratch for cardiology (your task), they just need specialized training on heart-related topics. The foundational knowledge <strong>transfers</strong>.
</div>
<h3>Two Strategies</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Feature Extraction (Freeze backbone)</strong><br>Freeze all pretrained layers. Only train a new classification head. Best when you have <strong>little data</strong> (<1000 images) or your task is similar to ImageNet.</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Fine-tuning (Unfreeze some layers)</strong><br>Unfreeze the last few layers and train with a very small learning rate (1e-4 to 1e-5). Best when you have <strong>moderate data</strong> (1000-50000 images) or your task is somewhat different.</div></div>
<h3>When to Use Which Strategy</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Data Size</th>
<th style="padding: 10px; color: var(--cyan);">Similar to ImageNet?</th>
<th style="padding: 10px; color: var(--cyan);">Strategy</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Small (<1K)</td><td>Yes</td><td>Feature extraction</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Small (<1K)</td><td>No</td><td>Feature extraction + heavy augmentation</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Medium (1K-50K)</td><td>Yes</td><td>Fine-tune last few layers</td></tr>
<tr><td style="padding: 8px;">Large (>50K)</td><td>No</td><td>Fine-tune entire network (low lr)</td></tr>
</table>
`,
concepts: `
<h3>Why Transfer Learning Works</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Feature Hierarchy:</strong> Early layers learn universal features (edges, textures) that transfer across domains</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Domain Similarity:</strong> The more similar source and target domains, the better transfer</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Regularization Effect:</strong> Pre-trained weights act as strong priors, preventing overfitting</div>
</div>
<h3>Transfer Learning Quadrant</h3>
<table>
<tr>
<th></th>
<th>Similar Domain</th>
<th>Different Domain</th>
</tr>
<tr>
<td><strong>Large Data</strong></td>
<td>Fine-tune all layers</td>
<td>Fine-tune top layers</td>
</tr>
<tr>
<td><strong>Small Data</strong></td>
<td>Feature extraction</td>
<td>Feature extraction (risky)</td>
</tr>
</table>
`,
math: `
<h3>Learning Rate Strategies</h3>
<p>Different layers need different learning rates during fine-tuning.</p>
<div class="formula">
Discriminative Fine-tuning:<br>
lr_layer_n = lr_base ร decay^(L-n)<br>
<br>
Where L = total layers, n = layer index<br>
Example: lr_base=1e-3, decay=0.9<br>
Layer 1: 1e-3 ร 0.9^9 โ 3.9e-4<br>
Layer 10: 1e-3 ร 0.9^0 = 1e-3
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Domain Shift</div>
When source and target distributions differ:<br>
โข <strong>Covariate Shift:</strong> P(X) changes, P(Y|X) same<br>
โข <strong>Label Shift:</strong> P(Y) changes, P(X|Y) same<br>
โข <strong>Concept Shift:</strong> P(Y|X) changes<br>
Transfer learning handles covariate shift well but struggles with concept shift.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฅ Medical Imaging</div>
<div class="box-content">
Train on ImageNet, fine-tune for X-ray diagnosis with only 1000 labeled images. Achieves 90%+ accuracy vs 60% from scratch.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Retail & E-commerce</div>
<div class="box-content">
Product classification, visual search, inventory management using pre-trained ResNet/EfficientNet models.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Satellite Imagery</div>
<div class="box-content">
Land use classification, deforestation detection, urban planning using models pre-trained on aerial imagery.
</div>
</div>
`,
code: `<span class="code-title">๐ transfer_learning.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torchvision.models <span class="keyword">as</span> models
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="comment"># --- Load pretrained ResNet-18 ---</span>
model = models.resnet18(weights=<span class="string">"IMAGENET1K_V1"</span>)
<span class="comment"># --- Strategy 1: Feature Extraction (freeze all) ---</span>
<span class="keyword">for</span> param <span class="keyword">in</span> model.parameters():
param.requires_grad = <span class="keyword">False</span>
<span class="comment"># Replace final layer for 5 classes</span>
model.fc = nn.Linear(model.fc.in_features, <span class="number">5</span>)
<span class="comment"># Only train the new layer</span>
optimizer = torch.optim.Adam(model.fc.parameters(), lr=<span class="number">0.001</span>)
<span class="comment"># --- Strategy 2: Fine-tuning (unfreeze some layers) ---</span>
<span class="keyword">for</span> param <span class="keyword">in</span> model.layer4.parameters():
param.requires_grad = <span class="keyword">True</span> <span class="comment"># unfreeze last residual block</span>
<span class="keyword">print</span>(f<span class="string">"Trainable: {sum(p.numel() for p in model.parameters() if p.requires_grad):,}"</span>)</div>
`
},
"localization": {
overview: `
<h3>Object Localization</h3>
<p>Predict both class and bounding box for a single object in image.</p>
<h3>Multi-Task Loss</h3>
<div class="formula">
Total Loss = L_classification + ฮป ร L_bbox<br>
<br>
Where:<br>
L_classification = Cross-Entropy<br>
L_bbox = Smooth L1 or IoU loss<br>
ฮป = balance term (typically 1-10)
</div>
<h3>Bounding Box Representation</h3>
<ul>
<li><strong>Option 1:</strong> (x_min, y_min, x_max, y_max)</li>
<li><strong>Option 2:</strong> (x_center, y_center, width, height) โ Most common</li>
</ul>
`,
concepts: `
<h3>Localization vs Detection</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Classification:</strong> What is in the image? โ "Cat"</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Localization:</strong> Where is the single object? โ "Cat at [100, 50, 200, 150]"</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Detection:</strong> Where are ALL objects? โ Multiple bounding boxes</div>
</div>
<h3>Network Architecture</h3>
<p>Modify a classification network (ResNet, VGG) by adding a regression head:</p>
<div class="formula">
CNN Backbone โ Feature Map โ [Classification Head (1000 classes)]<br>
โ [Regression Head (4 coordinates)]
</div>
`,
math: `
<h3>Smooth L1 Loss (Huber Loss)</h3>
<p>Combines L1 and L2 loss for robust bounding box regression.</p>
<div class="formula">
SmoothL1(x) = { 0.5xยฒ if |x| < 1<br>
{ |x| - 0.5 otherwise
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Why Smooth L1?</div>
โข <strong>L2 Loss:</strong> Penalizes large errors too much (squared), sensitive to outliers<br>
โข <strong>L1 Loss:</strong> Robust to outliers but has discontinuous gradient at 0<br>
โข <strong>Smooth L1:</strong> Best of both worlds - quadratic near 0, linear for large errors
</div>
<h3>IoU Loss</h3>
<div class="formula">
L_IoU = 1 - IoU(pred, target)<br>
Where IoU = Intersection / Union
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Self-Driving Cars</div>
<div class="box-content">Localize the primary vehicle ahead for adaptive cruise control</div>
</div>
<div class="info-box">
<div class="box-title">๐ธ Photo Auto-Crop</div>
<div class="box-content">Detect main subject and automatically crop to optimal composition</div>
</div>
<div class="info-box">
<div class="box-title">๐ฅ Medical Imaging</div>
<div class="box-content">Localize tumors, organs, or anomalies in X-rays and CT scans</div>
</div>
`
},
"rcnn": {
overview: `
<h3>R-CNN Family Evolution</h3>
<table>
<tr>
<th>Model</th>
<th>Year</th>
<th>Speed (FPS)</th>
<th>Key Innovation</th>
</tr>
<tr>
<td>R-CNN</td>
<td>2014</td>
<td>0.05</td>
<td>Selective Search + CNN features</td>
</tr>
<tr>
<td>Fast R-CNN</td>
<td>2015</td>
<td>0.5</td>
<td>RoI Pooling (share conv features)</td>
</tr>
<tr>
<td>Faster R-CNN</td>
<td>2015</td>
<td>7</td>
<td>Region Proposal Network (RPN)</td>
</tr>
<tr>
<td>Mask R-CNN</td>
<td>2017</td>
<td>5</td>
<td>+ Instance Segmentation masks</td>
</tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ก When to Use</div>
Faster R-CNN: Best accuracy for detection (not real-time)<br>
Mask R-CNN: Detection + instance segmentation
</div>
`,
concepts: `
<h3>Two-Stage Detection Pipeline</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Stage 1 - Region Proposal:</strong> Find ~2000 candidate regions that might contain objects</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Stage 2 - Classification:</strong> Classify each region and refine bounding box</div>
</div>
<h3>Region Proposal Network (RPN)</h3>
<p>The key innovation of Faster R-CNN - learns to propose regions instead of using hand-crafted algorithms.</p>
<div class="formula">
RPN Output per location:<br>
โข k anchor boxes ร 4 coordinates = 4k regression outputs<br>
โข k anchor boxes ร 2 objectness scores = 2k classification outputs<br>
Typical k = 9 (3 scales ร 3 aspect ratios)
</div>
`,
math: `
<h3>RoI Pooling: Fixed-Size Feature Maps</h3>
<p>Convert variable-size regions into fixed 7ร7 feature maps for FC layers.</p>
<div class="formula">
For each RoI of size HรW:<br>
1. Divide into 7ร7 grid (cells of size H/7 ร W/7)<br>
2. Max-pool each cell โ single value<br>
3. Output: Fixed 7ร7 feature map regardless of input size
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: RoI Align vs RoI Pool</div>
<strong>Problem:</strong> RoI Pooling quantizes coordinates, causing misalignment.<br>
<strong>Solution:</strong> RoI Align uses bilinear interpolation instead of rounding.<br>
This is critical for Mask R-CNN where pixel-level accuracy matters!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฅ Medical Imaging</div>
<div class="box-content">High-accuracy tumor detection where speed is less critical than precision</div>
</div>
<div class="info-box">
<div class="box-title">๐ท Photo Analysis</div>
<div class="box-content">Face detection, scene understanding, object counting in static images</div>
</div>
<div class="info-box">
<div class="box-title">๐ฌ Scientific Research</div>
<div class="box-content">Cell detection, particle tracking, microscopy image analysis</div>
</div>
`
},
"ssd": {
overview: `
<h3>SSD (Single Shot MultiBox Detector)</h3>
<p>Balances speed and accuracy by predicting boxes at multiple scales.</p>
<h3>Key Ideas</h3>
<ul>
<li><strong>Multi-Scale:</strong> Predictions from different layers (early = small objects, deep = large)</li>
<li><strong>Default Boxes (Anchors):</strong> Pre-defined boxes of various aspects ratios</li>
<li><strong>Single Pass:</strong> No separate region proposal step</li>
</ul>
<div class="callout insight">
<div class="callout-title">๐ Performance</div>
SSD300: 59 FPS, 74.3% mAP<br>
SSD512: 22 FPS, 76.8% mAP<br>
<br>
Sweet spot between YOLO (faster) and Faster R-CNN (more accurate)
</div>
`,
concepts: `
<h3>Multi-Scale Feature Maps</h3>
<p>SSD makes predictions at multiple layers, each detecting objects at different scales.</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Early Layers (38ร38):</strong> Detect small objects (high resolution)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Middle Layers (19ร19, 10ร10):</strong> Detect medium objects</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Deep Layers (5ร5, 3ร3, 1ร1):</strong> Detect large objects</div>
</div>
<h3>Default Boxes (Anchors)</h3>
<p>At each feature map cell, SSD predicts offsets for k default boxes with different aspect ratios (1:1, 2:1, 1:2, 3:1, 1:3).</p>
`,
math: `
<h3>SSD Loss Function</h3>
<p>Weighted sum of localization and confidence losses.</p>
<div class="formula">
L(x, c, l, g) = (1/N) ร [L_conf(x, c) + ฮฑ ร L_loc(x, l, g)]<br>
<br>
Where:<br>
โข L_conf = Softmax loss over class confidences<br>
โข L_loc = Smooth L1 loss over box coordinates<br>
โข ฮฑ = Weight factor (typically 1)<br>
โข N = Number of matched default boxes
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Hard Negative Mining</div>
Problem: Most default boxes are background (class imbalance).<br>
Solution: Sort negative boxes by confidence loss, pick top ones so pos:neg = 1:3.<br>
This focuses training on hard negatives, not easy ones.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐น Video Analytics</div>
<div class="box-content">Real-time object detection in security cameras, sports broadcasting</div>
</div>
<div class="info-box">
<div class="box-title">๐ค Robotics</div>
<div class="box-content">Object detection for manipulation tasks, obstacle avoidance</div>
</div>
<div class="info-box">
<div class="box-title">๐ฑ Mobile Apps</div>
<div class="box-content">Lightweight models for on-device detection (MobileNet-SSD)</div>
</div>
`
},
"semantic-seg": {
overview: `
<h3>Semantic Segmentation</h3>
<p>Classify every pixel in the image (pixel-wise classification).</p>
<h3>Popular Architectures</h3>
<table>
<tr>
<th>Model</th>
<th>Key Feature</th>
</tr>
<tr>
<td>FCN</td>
<td>Fully Convolutional (no FC layers)</td>
</tr>
<tr>
<td>U-Net</td>
<td>Skip connections from encoder to decoder</td>
</tr>
<tr>
<td>DeepLab</td>
<td>Atrous (dilated) convolutions + ASPP</td>
</tr>
</table>
<div class="formula">
U-Net Pattern:<br>
Input โ Encoder (downsample) โ Bottleneck โ Decoder (upsample) โ Pixel-wise Output<br>
With skip connections from encoder to decoder at each level
</div>
`,
concepts: `
<h3>Key Concepts</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Encoder-Decoder:</strong> Downsample to capture context, upsample to recover spatial detail</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Skip Connections:</strong> Pass high-resolution features from encoder to decoder (U-Net)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Atrous Convolution:</strong> Expand receptive field without losing resolution (DeepLab)</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>ASPP:</strong> Atrous Spatial Pyramid Pooling - capture multi-scale context</div>
</div>
`,
math: `
<h3>Dice Loss for Segmentation</h3>
<p>Better than cross-entropy for imbalanced classes (small objects).</p>
<div class="formula">
Dice = 2 ร |A โฉ B| / (|A| + |B|)<br>
Dice Loss = 1 - Dice<br>
<br>
Where A = predicted mask, B = ground truth mask
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Why Dice > Cross-Entropy?</div>
If only 1% of pixels are foreground:<br>
โข Cross-Entropy: Model can get 99% accuracy by predicting all background!<br>
โข Dice: Penalizes missed foreground pixels heavily<br>
โข Often use combination: L = BCE + Dice
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฅ Medical Imaging</div>
<div class="box-content">Tumor segmentation, organ delineation, cell analysis</div>
</div>
<div class="info-box">
<div class="box-title">๐ Autonomous Driving</div>
<div class="box-content">Road segmentation, free space detection, drivable area</div>
</div>
`
},
"instance-seg": {
overview: `
<h3>Instance Segmentation</h3>
<p>Detect AND segment each individual object (combines object detection + semantic segmentation).</p>
<h3>Difference from Semantic Segmentation</h3>
<ul>
<li><strong>Semantic:</strong> All "person" pixels get same label</li>
<li><strong>Instance:</strong> Person #1, Person #2, Person #3 (separate instances)</li>
</ul>
<h3>Main Approach: Mask R-CNN</h3>
<div class="formula">
Faster R-CNN + Segmentation Branch<br>
<br>
For each RoI:<br>
1. Bounding box regression<br>
2. Class prediction<br>
3. <strong>Binary mask for the object</strong>
</div>
`,
concepts: `
<h3>Mask R-CNN Architecture</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Backbone:</strong> ResNet-50/101 with Feature Pyramid Network (FPN)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>RPN:</strong> Region Proposal Network (same as Faster R-CNN)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>RoI Align:</strong> Better than RoI Pooling (no quantization)</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Mask Head:</strong> Small FCN that outputs 28ร28 binary mask per class</div>
</div>
`,
math: `
<h3>Multi-Task Loss</h3>
<p>Mask R-CNN optimizes three losses simultaneously:</p>
<div class="formula">
L = L_cls + L_box + L_mask<br>
<br>
Where:<br>
โข L_cls = Classification loss (cross-entropy)<br>
โข L_box = Bounding box regression (smooth L1)<br>
โข L_mask = Binary cross-entropy per-pixel mask loss
</div>
<div class="callout insight">
<div class="callout-title">๐ Key Insight: Decoupled Masks</div>
Mask R-CNN predicts a binary mask for EACH class independently.<br>
This avoids competition between classes and improves accuracy.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ธ Photo Editing</div>
<div class="box-content">Auto-select objects for editing, background removal, composition</div>
</div>
<div class="info-box">
<div class="box-title">๐ค Robotics</div>
<div class="box-content">Object manipulation - need exact shape, not just bounding box</div>
</div>
<div class="info-box">
<div class="box-title">๐ฌ Video Production</div>
<div class="box-content">Rotoscoping, VFX, green screen replacement</div>
</div>
`
},
"face-recog": {
overview: `
<h3>Face Recognition with Siamese Networks</h3>
<p>Learn similarity between faces using metric learning instead of classification.</p>
<h3>Triplet Loss Training</h3>
<div class="formula">
Loss = max(||f(A) - f(P)||ยฒ - ||f(A) - f(N)||ยฒ + margin, 0)<br>
<br>
Where:<br>
A = Anchor (reference face)<br>
P = Positive (same person)<br>
N = Negative (different person)<br>
margin = minimum separation (e.g., 0.2)
</div>
<div class="callout tip">
<div class="callout-title">๐ก One-Shot Learning</div>
After training, recognize new people with just 1-2 photos!<br>
No retraining needed - just compare embeddings.
</div>
`,
concepts: `
<h3>Face Recognition Pipeline</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Face Detection:</strong> Find faces in image (MTCNN, RetinaFace)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Alignment:</strong> Normalize face orientation and scale</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Embedding:</strong> Extract 128/512-dim feature vector (FaceNet, ArcFace)</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Matching:</strong> Compare embeddings with cosine similarity or L2 distance</div>
</div>
<h3>Key Models</h3>
<table>
<tr><th>Model</th><th>Key Innovation</th></tr>
<tr><td>FaceNet</td><td>Triplet loss, 128-dim embedding</td></tr>
<tr><td>ArcFace</td><td>Additive angular margin loss, SOTA accuracy</td></tr>
<tr><td>DeepFace</td><td>Facebook's early success</td></tr>
</table>
`,
math: `
<h3>Triplet Loss Intuition</h3>
<p>Push same-person faces closer, different-person faces apart.</p>
<div class="formula">
||f(A) - f(P)||ยฒ + margin < ||f(A) - f(N)||ยฒ
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Hard Triplet Mining</div>
Easy triplets: Random selection - margin already satisfied, loss=0<br>
Hard triplets: Find P closest to anchor, N closest to anchor from different class<br>
<strong>Training on hard triplets is critical for convergence!</strong>
</div>
<h3>ArcFace Angular Margin</h3>
<div class="formula">
L = -log(e^(sยทcos(ฮธ + m)) / (e^(sยทcos(ฮธ + m)) + ฮฃ e^(sยทcos(ฮธ_j))))<br>
Where m = angular margin, s = scale factor
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฑ Phone Unlock</div>
<div class="box-content">Face ID, biometric authentication</div>
</div>
<div class="info-box">
<div class="box-title">๐ Security</div>
<div class="box-content">Access control, surveillance, identity verification</div>
</div>
`
},
"autoencoders": {
overview: `
<h3>Autoencoders</h3>
<p>Unsupervised learning to compress data into latent representation and reconstruct it.</p>
<h3>Architecture</h3>
<div class="formula">
Input โ Encoder โ Latent Code (bottleneck) โ Decoder โ Reconstruction<br>
<br>
Loss = ||Input - Reconstruction||ยฒ (MSE)
</div>
<h3>Variants</h3>
<ul>
<li><strong>Vanilla:</strong> Basic autoencoder</li>
<li><strong>Denoising:</strong> Input corrupted, output clean (learns robust features)</li>
<li><strong>Variational (VAE):</strong> Probabilistic latent space (for generation)</li>
<li><strong>Sparse:</strong> Encourage sparse activations</li>
</ul>
`,
concepts: `
<h3>Key Concepts</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Bottleneck:</strong> Force information compression by using fewer dimensions than input</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Reconstruction:</strong> Learn to recreate input - captures essential features</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Latent Space:</strong> Compressed representation captures data structure</div>
</div>
<h3>Variational Autoencoder (VAE)</h3>
<p>Instead of encoding to a point, encode to a probability distribution (mean + variance).</p>
<div class="formula">
Encoder outputs: ฮผ (mean) and ฯ (standard deviation)<br>
Sample: z = ฮผ + ฯ ร ฮต (where ฮต ~ N(0,1))<br>
This is the "reparameterization trick" for backprop!
</div>
`,
math: `
<h3>VAE Loss Function (ELBO)</h3>
<p>VAE maximizes the Evidence Lower Bound:</p>
<div class="formula">
L = E[log p(x|z)] - KL(q(z|x) || p(z))<br>
<br>
Where:<br>
โข First term: Reconstruction quality<br>
โข Second term: KL divergence regularization (push q toward N(0,1))
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: KL Divergence</div>
For Gaussians:<br>
KL = -0.5 ร ฮฃ(1 + log(ฯยฒ) - ฮผยฒ - ฯยฒ)<br>
This has a closed-form solution - no sampling needed!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐๏ธ Compression</div>
<div class="box-content">Dimensionality reduction, data compression, feature extraction</div>
</div>
<div class="info-box">
<div class="box-title">๐ Anomaly Detection</div>
<div class="box-content">High reconstruction error = anomaly (fraud detection, defect detection)</div>
</div>
`
},
"gans": {
overview: `
<p>Generative Adversarial Networks (GANs) are a framework for training generative models using an <strong>adversarial game</strong> between two neural networks: a <strong>Generator</strong> (forger) and a <strong>Discriminator</strong> (detective).</p>
<div class="callout insight">
<div class="callout-title">๐จ The Art Forger Analogy</div>
The Generator is an art forger trying to create fake paintings that look real. The Discriminator is an art detective trying to spot fakes. As the detective gets better at spotting fakes, the forger improves to fool the detective. Eventually, the forger creates paintings so good that even the detective can't tell them apart from real ones.
</div>
<h3>The Min-Max Game</h3>
<div class="formula" style="text-align:center;">
$$\\min_G \\max_D \\; V(D,G) = \\mathbb{E}_{x \\sim p_{data}}[\\log D(x)] + \\mathbb{E}_{z \\sim p_z}[\\log(1 - D(G(z)))]$$
</div>
<p><strong>D(x)</strong> = probability that $x$ is real. <strong>G(z)</strong> = fake image generated from noise $z$.</p>
<h3>GAN Variants</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Variant</th>
<th style="padding: 10px; color: var(--cyan);">Innovation</th>
<th style="padding: 10px; color: var(--cyan);">Use Case</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">DCGAN</td><td>Convolutional architecture</td><td>Image generation</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">WGAN</td><td>Wasserstein distance (stable training)</td><td>Any generation task</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Pix2Pix</td><td>Paired image-to-image translation</td><td>Sketch โ Photo</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">CycleGAN</td><td>Unpaired image translation</td><td>Horse โ Zebra</td></tr>
<tr><td style="padding: 8px;">StyleGAN</td><td>Style-based generation (faces)</td><td>High-res face synthesis</td></tr>
</table>
<div class="callout warning">
<div class="callout-title">โ ๏ธ GAN Training Challenges</div>
<strong>Mode Collapse:</strong> Generator produces only one type of output. Solution: use WGAN or add diversity penalties.<br>
<strong>Training Instability:</strong> G and D can oscillate without converging. Solution: use spectral normalization, gradient penalty.<br>
<strong>Evaluation:</strong> Hard to objectively measure quality. Metrics like FID (Frรฉchet Inception Distance) help but aren't perfect.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐จ Image Generation</div>
<div class="box-content">
<strong>StyleGAN:</strong> Photorealistic faces, art generation<br>
<strong>DCGAN:</strong> Bedroom images, object generation
</div>
</div>
`,
math: `
<h3>The Minimax Game Objective</h3>
<p>The original GAN objective from Ian Goodfellow (2014) is a zero-sum game between Discriminator (D) and Generator (G).</p>
<div class="formula" style="font-size: 1.1rem; padding: 20px;">
min_G max_D V(D, G) = E_xโผp_data[log D(x)] + E_zโผp_z[log(1 - D(G(z)))]
</div>
<h3>Paper & Pain: Finding the Optimal Discriminator</h3>
<p>For a fixed Generator, the optimal Discriminator D* is:</p>
<div class="formula">
D*(x) = p_data(x) / (p_data(x) + p_g(x))
</div>
<div class="callout insight">
<div class="callout-title">๐ Theoretical Insight</div>
When the Discriminator is optimal, the Generator's task is essentially to minimize the <strong>Jensen-Shannon Divergence (JSD)</strong> between the data distribution and the model distribution. <br>
<strong>Problem:</strong> JSD is "flat" when distributions don't overlap, leading to vanishing gradients. This is why <strong>Wasserstein GAN (WGAN)</strong> was inventedโusing Earth Mover's distance instead!
</div>
<h3>Generator Gradient Problem</h3>
<p>Early in training, D(G(z)) is near 0. The term log(1-D(G(z))) has a very small gradient. </p>
<div class="list-item">
<div class="list-num">๐ก</div>
<div><strong>Heuristic Fix:</strong> Instead of minimizing log(1-D(G(z))), we maximize <strong>log D(G(z))</strong>. This provides much stronger gradients early on!</div>
</div>
`,
code: `
<h3>GAN Objective (Min-Max Game)</h3>
<div class="formula" style="text-align:center;">
$$\\min_G \\max_D \\; V(D,G) = \\mathbb{E}_{x \\sim p_{data}}[\\log D(x)] + \\mathbb{E}_{z \\sim p_z}[\\log(1 - D(G(z)))]$$
</div>
<p><strong>Generator</strong> tries to minimize โ create fakes so good that $D(G(z)) \\to 1$. <strong>Discriminator</strong> tries to maximize โ correctly classify real vs fake.</p>
<span class="code-title">๐ dcgan.py</span><div class="code-block"><span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">class</span> <span class="class-name">Generator</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self, latent_dim=<span class="number">100</span>):
<span class="builtin">super</span>().__init__()
self.net = nn.Sequential(
nn.ConvTranspose2d(latent_dim, <span class="number">512</span>, <span class="number">4</span>, <span class="number">1</span>, <span class="number">0</span>),
nn.BatchNorm2d(<span class="number">512</span>), nn.ReLU(),
nn.ConvTranspose2d(<span class="number">512</span>, <span class="number">256</span>, <span class="number">4</span>, <span class="number">2</span>, <span class="number">1</span>),
nn.BatchNorm2d(<span class="number">256</span>), nn.ReLU(),
nn.ConvTranspose2d(<span class="number">256</span>, <span class="number">128</span>, <span class="number">4</span>, <span class="number">2</span>, <span class="number">1</span>),
nn.BatchNorm2d(<span class="number">128</span>), nn.ReLU(),
nn.ConvTranspose2d(<span class="number">128</span>, <span class="number">3</span>, <span class="number">4</span>, <span class="number">2</span>, <span class="number">1</span>),
nn.Tanh(), <span class="comment"># Output: [-1, 1]</span>
)
<span class="keyword">def</span> <span class="function">forward</span>(self, z):
<span class="keyword">return</span> self.net(z.view(-<span class="number">1</span>, <span class="number">100</span>, <span class="number">1</span>, <span class="number">1</span>))
<span class="keyword">class</span> <span class="class-name">Discriminator</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self):
<span class="builtin">super</span>().__init__()
self.net = nn.Sequential(
nn.Conv2d(<span class="number">3</span>, <span class="number">64</span>, <span class="number">4</span>, <span class="number">2</span>, <span class="number">1</span>),
nn.LeakyReLU(<span class="number">0.2</span>),
nn.Conv2d(<span class="number">64</span>, <span class="number">128</span>, <span class="number">4</span>, <span class="number">2</span>, <span class="number">1</span>),
nn.BatchNorm2d(<span class="number">128</span>), nn.LeakyReLU(<span class="number">0.2</span>),
nn.Conv2d(<span class="number">128</span>, <span class="number">1</span>, <span class="number">4</span>, <span class="number">1</span>, <span class="number">0</span>),
nn.Sigmoid(),
)
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
<span class="keyword">return</span> self.net(x).view(-<span class="number">1</span>)</div>
`
},
"diffusion": {
overview: `
<h3>Diffusion Models</h3>
<p>Learn to reverse a gradual noising process, generating high-quality images.</p>
<h3>How Diffusion Works</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Forward Process:</strong> Gradually add Gaussian noise over T steps (xโ โ xโ โ ... โ x_T = pure noise)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Reverse Process:</strong> Train neural network to denoise (x_T โ x_{T-1} โ ... โ xโ = clean image)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Generation:</strong> Start from random noise, iteratively denoise T steps</div>
</div>
<div class="callout tip">
<div class="callout-title">โ
Advantages over GANs</div>
โข More stable training (no adversarial dynamics)<br>
โข Better sample quality and diversity<br>
โข Mode coverage (no mode collapse)<br>
โข Controllable generation (text-to-image)
</div>
`,
concepts: `
<h3>Key Components</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>U-Net Backbone:</strong> Encoder-decoder with skip connections predicts noise at each step</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Time Embedding:</strong> Tell the model which timestep it's at (sinusoidal encoding)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>CLIP Conditioning:</strong> Guide generation with text embeddings (Stable Diffusion)</div>
</div>
<h3>Latent Diffusion</h3>
<p>Instead of diffusing in pixel space (expensive), work in VAE latent space (8ร smaller).</p>
<div class="formula">
Image (512ร512ร3) โ VAE Encoder โ Latent (64ร64ร4) โ Diffuse โ Decode
</div>
`,
math: `
<h3>Forward Process (Noising)</h3>
<p>Add Gaussian noise according to a schedule ฮฒ_t:</p>
<div class="formula">
q(x_t | x_{t-1}) = N(x_t; โ(1-ฮฒ_t) ร x_{t-1}, ฮฒ_t ร I)<br>
<br>
Or in closed form for any t:<br>
x_t = โ(แพฑ_t) ร x_0 + โ(1-แพฑ_t) ร ฮต<br>
Where แพฑ_t = ฮ _{s=1}^t (1-ฮฒ_s)
</div>
<h3>Training Objective</h3>
<p>Simple noise prediction loss:</p>
<div class="formula">
L = E[||ฮต - ฮต_ฮธ(x_t, t)||ยฒ]
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Simplified Loss</div>
The full variational bound is complex, but Ho et al. (2020) showed this simple MSE loss on noise prediction works just as well and is much easier to implement!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ผ๏ธ Text-to-Image</div>
<div class="box-content">
<strong>Stable Diffusion:</strong> Open-source, runs on consumer GPUs<br>
<strong>DALL-E 2:</strong> OpenAI's photorealistic generator<br>
<strong>Midjourney:</strong> Artistic image generation
</div>
</div>
`
},
"rnn": {
overview: `
<p>Recurrent Neural Networks (RNNs) are designed for <strong>sequential data</strong> โ text, time series, audio, video frames. Unlike feedforward networks that process inputs independently, RNNs have a <strong>memory</strong> (hidden state) that carries information from previous time steps.</p>
<div class="callout insight">
<div class="callout-title">๐ The Reading Analogy</div>
When you read a sentence, you don't forget the beginning when you reach the end. Your brain maintains context. RNNs do the same โ they process one word (token) at a time while maintaining a hidden state that summarizes everything seen so far.
</div>
<h3>Vanilla RNN vs. LSTM vs. GRU</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Feature</th>
<th style="padding: 10px; color: var(--cyan);">Vanilla RNN</th>
<th style="padding: 10px; color: var(--cyan);">LSTM</th>
<th style="padding: 10px; color: var(--cyan);">GRU</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Memory</td><td>Short (5-10 steps)</td><td>Long (100+ steps)</td><td>Long</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Gates</td><td>None</td><td>3 (forget, input, output)</td><td>2 (reset, update)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Parameters</td><td>Fewest</td><td>Most (4x hidden_dimยฒ)</td><td>Middle (3x hidden_dimยฒ)</td></tr>
<tr><td style="padding: 8px;">Use Today?</td><td>โ Rarely</td><td>โ
Still useful</td><td>โ
Often preferred</td></tr>
</table>
<h3>The Vanishing Gradient Problem in RNNs</h3>
<p>Vanilla RNNs multiply the gradient by the same weight matrix at each time step. After many steps, the gradient either <strong>vanishes</strong> (matrix eigenvalues < 1) or <strong>explodes</strong> (eigenvalues > 1). LSTMs solve this with a special <strong>cell state</strong> highway that allows gradients to flow unchanged over many steps.</p>
<div class="callout tip">
<div class="callout-title">๐ฎ Historical Note</div>
RNNs dominated NLP from 2013-2017. The Transformer architecture (2017) has now largely replaced them because Transformers can process all tokens in parallel, while RNNs must process sequentially. However, RNNs are still used for real-time streaming data and edge devices.
</div>
`,
concepts: `
<h3>Step-by-Step LSTM Walkthrough</h3>
<p>LSTMs control the cell state via "Gates".</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Forget Gate layer:</strong> "What do we throw away?"<br>
<span class="formula-caption">ฯ(W_f ยท [h_{t-1}, x_t] + b_f)</span><br>
Output 0 = complete forget, 1 = keep everything.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Input Gate layer:</strong> "What new info do we add?"<br>
1. Sigmoid layer decides <strong>which values</strong> to update.<br>
2. Tanh layer creates <strong>new candidate values</strong> (~C_t).</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Update Cell State:</strong> The Critical Step<br>
<span class="formula" style="font-size:1.1rem">C_t = f_t * C_{t-1} + i_t * ~C_t</span><br>
We multiply old state by f_t (forgetting things) and add i_t * ~C_t (adding new things).</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Output Gate:</strong> "What do we reveal?"<br>
We filter the cell state: h_t = o_t * tanh(C_t).</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ก Intuition: Sigmoid vs Tanh</div>
โข <strong>Sigmoid (0 to 1):</strong> Acts like a valve or gate. Open/Close.<br>
โข <strong>Tanh (-1 to 1):</strong> Creates content/information. Normalized centered data.
</div>
`,
math: `
<h3>Paper & Pain: Why LSTMs Don't Vanish</h3>
<p>Let's look at the gradient flow in the Cell State equation:</p>
<div class="formula">
C_t = f_t \\cdot C_{t-1} + i_t \\cdot \\tilde{C}_t
</div>
<p>During backpropagation (BPTT), the derivative of C_t with respect to C_{t-1} is:</p>
<div class="formula" style="color: #00ff88;">
\\frac{\\partial C_t}{\\partial C_{t-1}} = f_t + \\dots
</div>
<div class="callout tip">
<div class="callout-title">โ
The Additive Gradient Highway</div>
In standard RNNs, we had multiplicative gradients (W^t). <br>
In LSTMs, the gradient is <strong>additive</strong>. If the Forget Gate f_t โ 1, the gradient passes through UNCHANGED (1.0).<br>
The error signal can travel back 1000 steps without vanishing!
</div>
<h3>Paper & Pain: Manual "Echo" Task</h3>
<p>Task: Input stream of 0s. If explicit "1" appears, remember it and output "1" 3 steps later.</p>
<div class="list-item">
<div class="list-num">๐</div>
<div><strong>Strategy:</strong> Set Input Gate to Open (1) only on trigger. Set Forget Gate to Closed (1) to maintain memory.</div>
</div>
<div class="list-item">
<div class="list-num">๐งฎ</div>
<div><strong>Weights:</strong> We can manually solve for weights that force i_t high only when x_t=1.</div>
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฃ๏ธ Sequence-to-Sequence</div>
<div class="box-content">
<strong>Translation:</strong> Google Translate (pre-Transformer) used massive LSTM stacks (GNMT).
</div>
</div>
<div class="info-box">
<div class="box-title">โ๏ธ Handwriting Generation</div>
<div class="box-content">
<strong>Alex Graves (2013):</strong> LSTMs can generate realistic cursive handwriting by predicting pen coordinates.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ต Music Composition</div>
<div class="box-content">
Generating melody and chords where context (key signature, tempo) must be maintained for minutes.
</div>
</div>
`,
code: `
<h3>LSTM Cell Equations</h3>
<div class="formula" style="text-align:center;">
<strong>Forget Gate:</strong> $$f_t = \\sigma(W_f \\cdot [h_{t-1}, x_t] + b_f)$$
<strong>Input Gate:</strong> $$i_t = \\sigma(W_i \\cdot [h_{t-1}, x_t] + b_i)$$
<strong>Cell Update:</strong> $$\\tilde{C}_t = \\tanh(W_C \\cdot [h_{t-1}, x_t] + b_C)$$
$$C_t = f_t \\odot C_{t-1} + i_t \\odot \\tilde{C}_t$$
<strong>Output Gate:</strong> $$o_t = \\sigma(W_o \\cdot [h_{t-1}, x_t] + b_o)$$
$$h_t = o_t \\odot \\tanh(C_t)$$
</div>
<span class="code-title">๐ lstm_sentiment.py</span><div class="code-block"><span class="keyword">import</span> torch
<span class="keyword">import</span> torch.nn <span class="keyword">as</span> nn
<span class="keyword">class</span> <span class="class-name">SentimentLSTM</span>(nn.Module):
<span class="keyword">def</span> <span class="function">__init__</span>(self, vocab_size, embed_dim, hidden_dim):
<span class="builtin">super</span>().__init__()
self.embedding = nn.Embedding(vocab_size, embed_dim)
self.lstm = nn.LSTM(embed_dim, hidden_dim, batch_first=<span class="keyword">True</span>, bidirectional=<span class="keyword">True</span>)
self.fc = nn.Linear(hidden_dim * <span class="number">2</span>, <span class="number">1</span>) <span class="comment"># *2 for bidirectional</span>
<span class="keyword">def</span> <span class="function">forward</span>(self, x):
embeds = self.embedding(x) <span class="comment"># (batch, seq_len, embed_dim)</span>
lstm_out, (h_n, c_n) = self.lstm(embeds) <span class="comment"># h_n: final hidden states</span>
hidden = torch.cat((h_n[-<span class="number">2</span>], h_n[-<span class="number">1</span>]), dim=<span class="number">1</span>)
<span class="keyword">return</span> torch.sigmoid(self.fc(hidden))
model = SentimentLSTM(vocab_size=<span class="number">10000</span>, embed_dim=<span class="number">128</span>, hidden_dim=<span class="number">256</span>)
x = torch.randint(<span class="number">0</span>, <span class="number">10000</span>, (<span class="number">4</span>, <span class="number">50</span>)) <span class="comment"># batch=4, seq_len=50</span>
<span class="keyword">print</span>(model(x).shape) <span class="comment"># (4, 1)</span></div>
`
},
"bert": {
overview: `
<p>BERT (Bidirectional Encoder Representations from Transformers) revolutionized NLP in 2018. Unlike GPT which reads left-to-right, BERT reads <strong>both directions simultaneously</strong>, giving it a deeper understanding of context.</p>
<div class="callout insight">
<div class="callout-title">๐ The Fill-in-the-Blank Analogy</div>
BERT was trained by randomly masking 15% of words and predicting them using the surrounding context (both left and right). It's like a student learning vocabulary by doing fill-in-the-blank exercises: "The cat sat on the [MASK]" โ "mat". This bidirectional context is what makes BERT so powerful for understanding language.
</div>
<h3>Two Training Objectives</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Masked Language Modeling (MLM):</strong> Randomly mask 15% of tokens and predict them. This forces the model to understand context from both sides.</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Next Sentence Prediction (NSP):</strong> Given two sentences, predict if sentence B follows sentence A. This teaches the model inter-sentence relationships. (Note: Later papers showed NSP isn't crucial.)</div></div>
<h3>BERT vs GPT</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Feature</th>
<th style="padding: 10px; color: var(--cyan);">BERT</th>
<th style="padding: 10px; color: var(--cyan);">GPT</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Direction</td><td>Bidirectional</td><td>Left-to-right (autoregressive)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Best for</td><td>Understanding (classification, QA, NER)</td><td>Generation (text, code, chat)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Architecture</td><td>Transformer Encoder</td><td>Transformer Decoder</td></tr>
<tr><td style="padding: 8px;">Parameters</td><td>110M (base) / 340M (large)</td><td>175B (GPT-3) / 1.7T (GPT-4)</td></tr>
</table>
`,
concepts: `
<h3>BERT Architecture</h3>
<p>BERT uses a multi-layer bidirectional Transformer encoder based on Vaswani et al. (2017).</p>
<h3>Model Variants</h3>
<table>
<tr>
<th>Model</th>
<th>Layers (L)</th>
<th>Hidden Size (H)</th>
<th>Attention Heads (A)</th>
<th>Parameters</th>
</tr>
<tr>
<td>BERT<sub>BASE</sub></td>
<td>12</td>
<td>768</td>
<td>12</td>
<td>110M</td>
</tr>
<tr>
<td>BERT<sub>LARGE</sub></td>
<td>24</td>
<td>1024</td>
<td>16</td>
<td>340M</td>
</tr>
</table>
<p><em>Note: BERT<sub>BASE</sub> was designed to match GPT's size for comparison.</em></p>
<h3>Input Representation</h3>
<p>BERT's input embedding is the sum of three components:</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Token Embeddings:</strong> WordPiece tokenization with 30,000 token vocabulary. Handles unknown words by splitting into subwords (e.g., "playing" โ "play" + "##ing")</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Segment Embeddings:</strong> Learned embedding to distinguish sentence A from sentence B (E<sub>A</sub> or E<sub>B</sub>)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Position Embeddings:</strong> Learned positional encodings (unlike Transformers' sinusoidal), supports sequences up to 512 tokens</div>
</div>
<div class="formula">
Input = Token_Embedding + Segment_Embedding + Position_Embedding
</div>
<h3>Special Tokens</h3>
<div class="info-box">
<div class="box-title">๐ท๏ธ Special Token Usage</div>
<div class="box-content">
<strong>[CLS]:</strong> Prepended to every input. Final hidden state used for classification tasks<br>
<strong>[SEP]:</strong> Separates sentence pairs and marks sequence end<br>
<strong>[MASK]:</strong> Replaces masked tokens during pre-training (not used during fine-tuning)
</div>
</div>
<h4>Example Input Format</h4>
<div class="formula">
[CLS] My dog is cute [SEP] He likes playing [SEP]<br>
<br>
Tokens: [CLS] My dog is cute [SEP] He likes play ##ing [SEP]<br>
Segments: E_A E_A E_A E_A E_A E_A E_B E_B E_B E_B E_B<br>
Positions: 0 1 2 3 4 5 6 7 8 9 10
</div>
<h3>Fine-tuning for Different Tasks</h3>
<table>
<tr><th>Task Type</th><th>Input Format</th><th>Output</th></tr>
<tr>
<td>Classification</td>
<td>[CLS] text [SEP]</td>
<td>[CLS] representation โ classifier</td>
</tr>
<tr>
<td>Sentence Pair</td>
<td>[CLS] sent A [SEP] sent B [SEP]</td>
<td>[CLS] representation โ classifier</td>
</tr>
<tr>
<td>Question Answering</td>
<td>[CLS] question [SEP] passage [SEP]</td>
<td>Start/End span vectors over passage tokens</td>
</tr>
<tr>
<td>Token Classification</td>
<td>[CLS] text [SEP]</td>
<td>Each token representation โ label</td>
</tr>
</table>
`,
math: `
<h3>Pre-training Objective</h3>
<p>BERT simultaneously optimizes two unsupervised tasks:</p>
<div class="formula">
L = L<sub>MLM</sub> + L<sub>NSP</sub>
</div>
<h3>Masked Language Modeling (MLM)</h3>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: The Masking Strategy</div>
<strong>Problem:</strong> Standard left-to-right language modeling can't capture bidirectional context.<br>
<strong>Solution:</strong> Randomly mask 15% of tokens and predict them using full context.<br><br>
<strong>However:</strong> [MASK] token doesn't appear during fine-tuning!<br>
<strong>Clever Fix:</strong> Of the 15% selected tokens:<br>
โข 80% โ Replace with [MASK]<br>
โข 10% โ Replace with random token<br>
โข 10% โ Keep unchanged<br><br>
This forces the model to maintain context representations for ALL tokens!
</div>
<h4>MLM Loss Derivation</h4>
<p>Let's work through the MLM objective step by step:</p>
<div class="formula">
Given input sequence: x = [xโ, xโ, ..., x_n]<br>
Masked sequence: xฬ = [xฬโ, xฬโ, ..., xฬ_n]<br>
<br>
Let M = {iโ, iโ, ..., i_m} be indices of masked tokens<br>
<br>
For each masked position i โ M:<br>
h_i = BERT(xฬ)_i (hidden state at position i)<br>
logits_i = W ยท h_i + b (W โ โ^(VรH), vocab size V)<br>
P(x_i | xฬ) = softmax(logits_i)<br>
<br>
Cross-entropy loss per token:<br>
L_i = -log P(x_i | xฬ)<br>
<br>
Total MLM loss:<br>
L_MLM = (1/|M|) ฮฃ_{iโM} L_i<br>
L_MLM = -(1/|M|) ฮฃ_{iโM} log P(x_i | xฬ)
</div>
<div class="callout warning">
<div class="callout-title">๐ Worked Example: MLM Calculation</div>
<strong>Input:</strong> "The cat sat on the mat"<br>
<strong>After masking (15%):</strong> "The [MASK] sat on the mat"<br>
<strong>Target:</strong> Predict "cat" at position 2<br><br>
<strong>Step 1:</strong> Forward pass through BERT<br>
hโ = BERT(xฬ)โ โ โ^768 (for BERT_BASE)<br><br>
<strong>Step 2:</strong> Project to vocabulary space<br>
logitsโ = W ยท hโ + b โ โ^30000<br><br>
<strong>Step 3:</strong> Compute probabilities<br>
P(w | xฬ) = exp(logitsโ[w]) / ฮฃ_v exp(logitsโ[v])<br><br>
<strong>Step 4:</strong> Compute loss (assume P("cat"|xฬ) = 0.73)<br>
L = -log(0.73) = 0.315
</div>
<h3>Next Sentence Prediction (NSP)</h3>
<p>Binary classification task to understand sentence relationships.</p>
<div class="formula">
Input: [CLS] sentence_A [SEP] sentence_B [SEP]<br>
<br>
Let C = final hidden state of [CLS] token โ โ^H<br>
<br>
P(IsNext = True) = ฯ(W_NSP ยท C)<br>
where ฯ = sigmoid function, W_NSP โ โ^(1รH)<br>
<br>
Binary cross-entropy loss:<br>
L_NSP = -[yยทlog(ลท) + (1-y)ยทlog(1-ลท)]<br>
where y = 1 if B follows A, else 0
</div>
<h4>NSP Training Data Generation</h4>
<ul>
<li><strong>50% IsNext:</strong> B actually follows A in corpus</li>
<li><strong>50% NotNext:</strong> B sampled randomly from another document</li>
</ul>
<h3>Fine-tuning Math: Question Answering (SQuAD)</h3>
<div class="formula">
Input: [CLS] question [SEP] paragraph [SEP]<br>
<br>
Let T_i = final hidden state for token i in paragraph<br>
<br>
Start position logits: S_i = W_start ยท T_i<br>
End position logits: E_i = W_end ยท T_i<br>
<br>
P(start = i) = softmax(S)_i<br>
P(end = j) = softmax(E)_j<br>
<br>
Answer span = tokens from position i to j<br>
<br>
Training loss:<br>
L = -log P(start = i*) - log P(end = j*)<br>
where i*, j* are ground truth positions
</div>
`,
applications: `
<h3>SQuAD Benchmark Performance</h3>
<div class="info-box">
<div class="box-title">๐ Stanford Question Answering Dataset (SQuAD)</div>
<div class="box-content">
<strong>SQuAD 1.1:</strong> 100,000+ question-answer pairs on 500+ Wikipedia articles. Every question has an answer span in the passage.<br><br>
<strong>SQuAD 2.0:</strong> Adds 50,000+ unanswerable questions. Models must determine when no answer exists.<br><br>
<strong>Evaluation Metrics:</strong><br>
โข <strong>EM (Exact Match):</strong> % of predictions matching ground truth exactly<br>
โข <strong>F1:</strong> Token-level overlap between prediction and ground truth
</div>
</div>
<h3>SQuAD 1.1 Results</h3>
<table>
<tr><th>Model</th><th>EM</th><th>F1</th></tr>
<tr><td>Human Performance</td><td>82.3</td><td>91.2</td></tr>
<tr><td>BERT<sub>BASE</sub></td><td>80.8</td><td>88.5</td></tr>
<tr><td>BERT<sub>LARGE</sub></td><td><strong>84.1</strong></td><td><strong>90.9</strong></td></tr>
</table>
<p><em>BERT<sub>LARGE</sub> surpassed human performance on EM!</em></p>
<h3>SQuAD 2.0 Results</h3>
<table>
<tr><th>Model</th><th>EM</th><th>F1</th></tr>
<tr><td>Human Performance</td><td>86.9</td><td>89.5</td></tr>
<tr><td>BERT<sub>BASE</sub></td><td>73.7</td><td>76.3</td></tr>
<tr><td>BERT<sub>LARGE</sub></td><td><strong>78.7</strong></td><td><strong>81.9</strong></td></tr>
</table>
<div class="callout tip">
<div class="callout-title">๐ก Example SQuAD Question</div>
<strong>Passage:</strong> "The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France."<br><br>
<strong>Question:</strong> "In what country is Normandy located?"<br><br>
<strong>BERT Answer:</strong> "France" โ<br>
<strong>Start Token:</strong> position 32<br>
<strong>End Token:</strong> position 32
</div>
<h3>GLUE Benchmark (General Language Understanding Evaluation)</h3>
<p>BERT set new state-of-the-art on all 9 GLUE tasks:</p>
<table>
<tr><th>Task</th><th>Metric</th><th>Previous SOTA</th><th>BERT<sub>LARGE</sub></th></tr>
<tr><td>MNLI (NLI)</td><td>Acc</td><td>86.6</td><td><strong>86.7</strong></td></tr>
<tr><td>QQP (Paraphrase)</td><td>F1</td><td>66.1</td><td><strong>72.1</strong></td></tr>
<tr><td>QNLI (QA/NLI)</td><td>Acc</td><td>87.4</td><td><strong>92.7</strong></td></tr>
<tr><td>SST-2 (Sentiment)</td><td>Acc</td><td>93.2</td><td><strong>94.9</strong></td></tr>
<tr><td>CoLA (Acceptability)</td><td>Matthew's</td><td>35.0</td><td><strong>60.5</strong></td></tr>
</table>
<h3>Additional Applications</h3>
<div class="info-box">
<div class="box-title">๐ Google Search</div>
<div class="box-content">
In October 2019, Google began using BERT for 1 in 10 English search queries, calling it the biggest leap in 5 years. BERT helps understand search intent and context.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ท๏ธ Named Entity Recognition (NER)</div>
<div class="box-content">
BERT excels at identifying entities (person, location, organization) in text by treating it as token classification. Each token gets a label (B-PER, I-PER, B-LOC, etc.).
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Text Classification</div>
<div class="box-content">
Sentiment analysis, topic classification, spam detection - all benefit from BERT's contextual understanding. Simply use [CLS] representation with a classifier.
</div>
</div>
<h3>Using BERT: Quick Code Example</h3>
<div class="formula">
# Using Hugging Face Transformers<br>
from transformers import BertTokenizer, BertForQuestionAnswering<br>
import torch<br>
<br>
# Load pre-trained model and tokenizer<br>
tokenizer = BertTokenizer.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')<br>
model = BertForQuestionAnswering.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')<br>
<br>
# Example<br>
question = "What is BERT?"<br>
context = "BERT is a bidirectional Transformer for NLP."<br>
<br>
# Tokenize and get answer<br>
inputs = tokenizer(question, context, return_tensors='pt')<br>
outputs = model(**inputs)<br>
<br>
start_idx = torch.argmax(outputs.start_logits)<br>
end_idx = torch.argmax(outputs.end_logits)<br>
answer = tokenizer.convert_tokens_to_string(<br>
tokenizer.convert_ids_to_tokens(inputs['input_ids'][0][start_idx:end_idx+1])<br>
)<br>
print(answer) # "a bidirectional Transformer for NLP"
</div>
`,
code: `
<h3>BERT Training Objectives</h3>
<div class="formula" style="text-align:center;">
<strong>MLM:</strong> $$L_{MLM} = -\\sum_{i \\in M} \\log P(x_i | x_{\\backslash M})$$
<strong>NSP:</strong> $$L_{NSP} = -[y\\log(p) + (1-y)\\log(1-p)]$$
</div>
<span class="code-title">๐ bert_huggingface.py</span><div class="code-block"><span class="keyword">from</span> transformers <span class="keyword">import</span> BertTokenizer, BertModel, pipeline
<span class="comment"># --- Quick sentiment analysis ---</span>
classifier = pipeline(<span class="string">"sentiment-analysis"</span>)
result = classifier(<span class="string">"Deep learning is fascinating!"</span>)
<span class="keyword">print</span>(result) <span class="comment"># [{'label': 'POSITIVE', 'score': 0.9998}]</span>
<span class="comment"># --- BERT Embeddings for custom tasks ---</span>
tokenizer = BertTokenizer.from_pretrained(<span class="string">"bert-base-uncased"</span>)
model = BertModel.from_pretrained(<span class="string">"bert-base-uncased"</span>)
text = <span class="string">"Neural networks learn features automatically"</span>
inputs = tokenizer(text, return_tensors=<span class="string">"pt"</span>)
outputs = model(**inputs)
<span class="comment"># CLS token embedding (sentence representation)</span>
cls_embedding = outputs.last_hidden_state[:, <span class="number">0</span>, :]
<span class="keyword">print</span>(f<span class="string">"Shape: {cls_embedding.shape}"</span>) <span class="comment"># (1, 768)</span></div>
`
},
"gpt": {
overview: `
<p>GPT (Generative Pre-trained Transformer) is the architecture behind ChatGPT, GitHub Copilot, and the AI revolution. It's a <strong>decoder-only Transformer</strong> trained to predict the next token in a sequence โ a simple objective that produces remarkably intelligent behavior.</p>
<div class="callout insight">
<div class="callout-title">๐ฏ The Core Idea</div>
GPT's entire training objective is <strong>next-token prediction</strong>: given "The cat sat on the", predict "mat". By doing this on trillions of tokens from the internet, the model learns grammar, facts, reasoning, coding, math โ all as a byproduct of predicting what comes next. It's arguably the most powerful simple idea in AI.
</div>
<h3>The GPT Family</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Model</th>
<th style="padding: 10px; color: var(--cyan);">Year</th>
<th style="padding: 10px; color: var(--cyan);">Parameters</th>
<th style="padding: 10px; color: var(--cyan);">Key Innovation</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">GPT-1</td><td>2018</td><td>117M</td><td>Pretrain on unlabeled text, fine-tune</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">GPT-2</td><td>2019</td><td>1.5B</td><td>Zero-shot task transfer</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">GPT-3</td><td>2020</td><td>175B</td><td>In-context learning, few-shot prompting</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">GPT-4</td><td>2023</td><td>~1.7T (MoE)</td><td>Multimodal (text + images), RLHF</td></tr>
<tr><td style="padding: 8px;">GPT-4o</td><td>2024</td><td>Unknown</td><td>Native multimodal (audio, video, text)</td></tr>
</table>
<h3>How GPT Generates Text</h3>
<p>GPT uses <strong>autoregressive generation</strong>: it generates one token at a time, each conditioned on all previous tokens. The process:</p>
<div class="list-item"><div class="list-num">1</div><div>Feed prompt tokens through the Transformer decoder</div></div>
<div class="list-item"><div class="list-num">2</div><div>Get probability distribution over vocabulary (~50K tokens)</div></div>
<div class="list-item"><div class="list-num">3</div><div>Sample from the distribution (temperature, top-k, top-p control randomness)</div></div>
<div class="list-item"><div class="list-num">4</div><div>Append the new token and repeat from step 1</div></div>
`,
concepts: `
<h3>GPT Architecture</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Decoder Only:</strong> Uses causal (masked) attention - can only see past tokens</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Autoregressive:</strong> Generate one token at a time, feed back as input</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Pre-training:</strong> Next token prediction on massive text corpus</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>RLHF:</strong> Reinforcement Learning from Human Feedback (ChatGPT)</div>
</div>
<h3>In-Context Learning</h3>
<p>GPT-3+ can learn from examples in the prompt without updating weights!</p>
<div class="formula">
Zero-shot: "Translate to French: Hello" โ "Bonjour"<br>
Few-shot: "catโchat, dogโchien, houseโ?" โ "maison"
</div>
`,
math: `
<h3>Causal Language Modeling</h3>
<p>GPT is trained to maximize the likelihood of the next token:</p>
<div class="formula">
L = -ฮฃ log P(x_t | x_{<t})<br>
<br>
Where P(x_t | x_{<t}) = softmax(h_t ร W_vocab)
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Scaling Laws</div>
Performance scales predictably with compute, data, and parameters:<br>
L โ N^(-0.076) for model size N<br>
This is why OpenAI trained GPT-3 (175B) and GPT-4 (1.8T)!
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฌ ChatGPT & Assistants</div>
<div class="box-content">
Conversational AI, customer support, tutoring, brainstorming
</div>
</div>
<div class="info-box">
<div class="box-title">๐ป Code Generation</div>
<div class="box-content">
GitHub Copilot, code completion, bug fixing, documentation
</div>
</div>
`,
code: `<span class="code-title">๐ gpt_generation.py</span><div class="code-block"><span class="keyword">from</span> transformers <span class="keyword">import</span> GPT2LMHeadModel, GPT2Tokenizer
tokenizer = GPT2Tokenizer.from_pretrained(<span class="string">"gpt2"</span>)
model = GPT2LMHeadModel.from_pretrained(<span class="string">"gpt2"</span>)
<span class="comment"># Generate text</span>
prompt = <span class="string">"Deep learning has revolutionized"</span>
inputs = tokenizer.encode(prompt, return_tensors=<span class="string">"pt"</span>)
outputs = model.generate(
inputs, max_length=<span class="number">100</span>,
temperature=<span class="number">0.7</span>, <span class="comment"># creativity control</span>
top_k=<span class="number">50</span>, <span class="comment"># top-k sampling</span>
top_p=<span class="number">0.9</span>, <span class="comment"># nucleus sampling</span>
do_sample=<span class="keyword">True</span>,
no_repeat_ngram_size=<span class="number">2</span>,
)
text = tokenizer.decode(outputs[<span class="number">0</span>])
<span class="keyword">print</span>(text)</div>
`
},
"vit": {
overview: `
<h3>Vision Transformer (ViT)</h3>
<p>Apply Transformer architecture directly to images by treating them as sequences of patches.</p>
<h3>How ViT Works</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Patchify:</strong> Split 224ร224 image into 16ร16 patches (14ร14 = 196 patches)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Linear Projection:</strong> Flatten each patch โ linear embedding (like word embeddings)</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Positional Encoding:</strong> Add position information</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Transformer Encoder:</strong> Standard Transformer (self-attention, FFN)</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Classification:</strong> Use [CLS] token for final prediction</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ก When ViT Shines</div>
โข <strong>Large Datasets:</strong> Needs 10M+ images (or pre-training on ImageNet-21K)<br>
โข <strong>Transfer Learning:</strong> Pre-trained ViT beats CNNs on many tasks<br>
โข <strong>Long-Range Dependencies:</strong> Global attention vs CNN's local receptive field
</div>
`,
concepts: `
<h3>ViT vs CNN Comparison</h3>
<table>
<tr><th>Aspect</th><th>CNN</th><th>ViT</th></tr>
<tr><td>Inductive Bias</td><td>Locality, translation invariance</td><td>Minimal - learns from data</td></tr>
<tr><td>Data Efficiency</td><td>Better with small datasets</td><td>Needs large datasets</td></tr>
<tr><td>Receptive Field</td><td>Local (grows with depth)</td><td>Global from layer 1</td></tr>
<tr><td>Scalability</td><td>Diminishing returns</td><td>Scales well with compute</td></tr>
</table>
<h3>Key Innovations</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>No Convolutions:</strong> Pure attention - "An Image is Worth 16x16 Words"</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Learnable Position:</strong> Position embeddings are learned, not sinusoidal</div>
</div>
`,
math: `
<h3>Patch Embedding</h3>
<p>Convert image patches to token embeddings:</p>
<div class="formula">
z_0 = [x_cls; x_p^1 E; x_p^2 E; ...; x_p^N E] + E_pos<br>
<br>
Where:<br>
โข x_p^i = flattened patch (16ร16ร3 = 768 dimensions)<br>
โข E = learnable linear projection<br>
โข E_pos = position embedding
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: Computation</div>
ViT-Base: 12 layers, 768 hidden, 12 heads ~ 86M params<br>
Self-attention cost: O(nยฒยทd) where n=196 patches<br>
This is why ViT is efficient for images (196 tokens) vs text (1000+ tokens)
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ผ๏ธ Image Classification</div>
<div class="box-content">SOTA on ImageNet with pre-training. Google/DeepMind use for internal systems.</div>
</div>
<div class="info-box">
<div class="box-title">๐ Object Detection</div>
<div class="box-content">DETR, DINO - Transformer-based detection replacing Faster R-CNN</div>
</div>
<div class="info-box">
<div class="box-title">๐ฌ Video Understanding</div>
<div class="box-content">VideoViT, TimeSformer - extend patches to 3D (space + time)</div>
</div>
`
},
"seq2seq": {
overview: `
<h3>Seq2Seq with Attention</h3>
<p>The architecture that revolutionized Machine Translation (before Transformers).</p>
<h3>The Bottleneck Problem</h3>
<p>Standard Encoder-Decoder models try to compress the entire sentence "I love deep learning" into a single vector (Context Vector).</p>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Information Bottleneck</div>
For long sentences (e.g., 50 words), the fixed-size vector forgets early details. Performance degrades rapidly with length.
</div>
<h3>The Attention Solution</h3>
<p><strong>Idea:</strong> Don't just look at the last state. Let the Decoder "look back" at ALL Encoder states at every step.</p>
`,
concepts: `
<h3>Visualizing Attention (Alammar Style)</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Encoder States:</strong> We keep all hidden states $h_1, h_2, h_3, h_4$.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Alignment Scores:</strong> Decoder asks: "How relevant is $h_i$ to what I'm translating now?"</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Context Vector:</strong> Weighted sum of relevant states. If attention is strong on $h_2$, the context vector looks a lot like $h_2$.</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ก The "Searchlight" Analogy</div>
Attention is like a searchlight. When generating "รฉtudiant" (student), the light shines brightly on "student" in the input sentence.
</div>
`,
math: `
<h3>Attention Math (Bahdanau / Luong)</h3>
<p>Let Decoder state be $s_{t-1}$ and Encoder states be $h_j$.</p>
<div class="formula">
1. Score: $e_{tj} = score(s_{t-1}, h_j)$ (e.g., Dot Product)<br>
2. Weights: $\\alpha_{tj} = softmax(e_{tj})$<br>
3. Context: $c_t = \\sum \\alpha_{tj} h_j$<br>
4. Output: $s_t = RNN(s_{t-1}, [y_{t-1}, c_t])$
</div>
<div class="callout tip">
<div class="callout-title">๐ Paper & Pain: Dot Product vs Additive</div>
โข <strong>Dot Product (Luong):</strong> $s^T h$ (Fast, matrix mult)<br>
โข <strong>Additive (Bahdanau):</strong> $v^T tanh(W_1 s + W_2 h)$ (More parameters, originally better for large dim)
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ฃ๏ธ Neural Machine Translation</div>
<div class="box-content">Google Translate (2016) switched to GNMT (Attention-based), reducing errors by 60%.</div>
</div>
<div class="info-box">
<div class="box-title">๐ Text Summarization</div>
<div class="box-content">Focusing on key sentences in a long document to generate a headline.</div>
</div>
`
},
"research-papers": {
overview: `
<h3>Seminal Papers Library</h3>
<p>A curated collection of the most impactful papers in Deep Learning history, with "Paper & Pain" insights.</p>
<div class="callout insight">
<div class="callout-title">๐ How to Read Papers</div>
Don't just read the abstract. Look for the <strong>Objective Function</strong> and the <strong>Architecture Diagram</strong>. That's where the truth lies.
</div>
`,
concepts: `
<h3>Computer Vision Hall of Fame</h3>
<div class="list-item">
<div class="list-num">2012</div>
<div><strong>AlexNet</strong> (Krizhevsky et al.)<br>
<em>"ImageNet Classification with Deep Convolutional Neural Networks"</em><br>
<span class="formula-caption">Insight: Relied on TWO GPUs because 3GB RAM wasn't enough. The split architecture was distinct.</span><br>
<a href="https://proceedings.neurips.cc/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf" target="_blank" style="color: #ff6b35;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2014</div>
<div><strong>VGGNet</strong> (Simonyan & Zisserman)<br>
<em>"Very Deep Convolutional Networks for Large-Scale Image Recognition"</em><br>
<span class="formula-caption">Insight: 3x3 filters are all you need. Two 3x3 layers have the same receptive field as one 5x5 but fewer parameters.</span><br>
<a href="https://arxiv.org/pdf/1409.1556.pdf" target="_blank" style="color: #ff6b35;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2015</div>
<div><strong>U-Net</strong> (Ronneberger et al.)<br>
<em>"Convolutional Networks for Biomedical Image Segmentation"</em><br>
<span class="formula-caption">Insight: Skip connections concatenating features from encoder to decoder allow precise localization.</span><br>
<a href="https://arxiv.org/pdf/1505.04597.pdf" target="_blank" style="color: #ff6b35;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2015</div>
<div><strong>ResNet</strong> (He et al.)<br>
<em>"Deep Residual Learning for Image Recognition"</em><br>
<span class="formula-caption">Insight: It's easier to learn 0 than Identity. $f(x) = H(x) - x$.</span><br>
<a href="https://arxiv.org/pdf/1512.03385.pdf" target="_blank" style="color: #ff6b35;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2016</div>
<div><strong>YOLO</strong> (Redmon et al.)<br>
<em>"You Only Look Once: Unified, Real-Time Object Detection"</em><br>
<span class="formula-caption">Insight: Treated detection as a <strong>regression</strong> problem, not classification. Single forward pass.</span><br>
<a href="https://arxiv.org/pdf/1506.02640.pdf" target="_blank" style="color: #ff6b35;">๐ Read PDF</a></div>
</div>
<hr style="border-color: #333; margin: 20px 0;">
<h3>NLP & GenAI Hall of Fame</h3>
<div class="list-item">
<div class="list-num">2014</div>
<div><strong>GANs</strong> (Goodfellow et al.)<br>
<em>"Generative Adversarial Networks"</em><br>
<span class="formula-caption">Insight: Training a generator by fighting a discriminator. The minimax game: $\\min_G \\max_D V(D, G)$.</span><br>
<a href="https://arxiv.org/pdf/1406.2661.pdf" target="_blank" style="color: #a371f7;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2017</div>
<div><strong>Attention Is All You Need</strong> (Vaswani et al.)<br>
<em>"Transformers"</em><br>
<span class="formula-caption">Insight: Sinusoidal Positional Embeddings allow the model to generalize to lengths unseen during training.</span><br>
<a href="https://arxiv.org/pdf/1706.03762.pdf" target="_blank" style="color: #00d4ff;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2018</div>
<div><strong>BERT</strong> (Devlin et al.)<br>
<em>"Pre-training of Deep Bidirectional Transformers"</em><br>
<span class="formula-caption">Insight: Masked LM (Cloze task) is inefficient (only 15% signal) but crucial for bidirectionality.</span><br>
<a href="https://arxiv.org/pdf/1810.04805.pdf" target="_blank" style="color: #00d4ff;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2020</div>
<div><strong>GPT-3</strong> (Brown et al.)<br>
<em>"Language Models are Few-Shot Learners"</em><br>
<span class="formula-caption">Insight: Scale is all you need. 175B parameters enable emergent behavior like in-context learning.</span><br>
<a href="https://arxiv.org/pdf/2005.14165.pdf" target="_blank" style="color: #00d4ff;">๐ Read PDF</a></div>
</div>
<div class="list-item">
<div class="list-num">2020</div>
<div><strong>DDPM</strong> (Ho et al.)<br>
<em>"Denoising Diffusion Probabilistic Models"</em><br>
<span class="formula-caption">Insight: Predicting the noise $\\epsilon$ is mathematically equivalent to predicting the score function (gradient of data density).</span><br>
<a href="https://arxiv.org/pdf/2006.11239.pdf" target="_blank" style="color: #a371f7;">๐ Read PDF</a></div>
</div>
`,
math: `
<h3>The Formulas That Changed AI</h3>
<p><strong>ResNet Residual:</strong></p>
<div class="formula">y = F(x, \\{W_i\}) + x</div>
<p><strong>Scaled Dot-Product Attention:</strong></p>
<div class="formula">Attention(Q, K, V) = softmax(\\frac{QK^T}{\\sqrt{d_k}})V</div>
<p><strong>Diffusion Reverse Process:</strong></p>
<div class="formula">p_\\theta(x_{t-1}|x_t) = \\mathcal{N}(x_{t-1}; \\mu_\\theta(x_t, t), \\Sigma_\\theta(x_t, t))</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Impact</div>
<div class="box-content">These papers form the foundation of ChatGPT, Midjourney, Self-Driving Cars, and Facial Recognition.</div>
</div>
`
},
"gnn": {
overview: `
<h3>Graph Neural Networks (GNNs)</h3>
<p>Deep learning on non-Euclidean data structures like social networks, molecules, and knowledge graphs.</p>
<h3>Key Concepts</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Graph Structure:</strong> Nodes (entities) and Edges (relationships).</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Message Passing:</strong> Nodes exchange information with neighbors.</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Aggregation:</strong> Combine incoming messages (Sum, Mean, Max).</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ก Why GNNs?</div>
Standard CNNs expect a fixed grid (euclidean). Graphs have arbitrary size and topology. GNNs are permutation invariant!
</div>
`,
concepts: `
<h3>Message Passing Neural Networks (MPNN)</h3>
<p>The core framework for most GNNs.</p>
<div class="list-item">
<div class="list-num">1</div>
<div><strong>Message Function:</strong> Compute message from neighbor to node.</div>
</div>
<div class="list-item">
<div class="list-num">2</div>
<div><strong>Aggregation Function:</strong> Sum all messages from neighbors.</div>
</div>
<div class="list-item">
<div class="list-num">3</div>
<div><strong>Update Function:</strong> Update node state based on aggregated messages.</div>
</div>
`,
math: `
<h3>Graph Convolution Network (GCN)</h3>
<p>The "Hello World" of GNNs (Kipf & Welling, 2017).</p>
<div class="formula">
H^{(l+1)} = ฯ(D^{-1/2} A D^{-1/2} H^{(l)} W^{(l)})
</div>
<p>Where:</p>
<ul>
<li><strong>A:</strong> Adjacency Matrix (connections)</li>
<li><strong>D:</strong> Degree Matrix (number of connections)</li>
<li><strong>H:</strong> Node Features</li>
<li><strong>W:</strong> Learnable Weights</li>
</ul>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Over-smoothing</div>
If GNN is too deep, all node representations become indistinguishable. Usually 2-4 layers are enough.
</div>
`,
applications: `
<div class="info-box">
<div class="box-title">๐ Drug Discovery</div>
<div class="box-content">Predicting molecular properties, protein folding (AlphaFold)</div>
</div>
<div class="info-box">
<div class="box-title">๐ Traffic Prediction</div>
<div class="box-content">Road networks, estimating travel times (Google Maps)</div>
</div>
<div class="info-box">
<div class="box-title">๐ Recommender Systems</div>
<div class="box-content">Pinterest (PinSage), User-Item graphs</div>
</div>
`
},
"vector-db": {
overview: `
<p>Vector databases store and search <strong>high-dimensional embeddings</strong> โ the numerical representations that AI models use to understand text, images, and audio. They enable <strong>semantic search</strong>: finding items by meaning, not just keywords.</p>
<div class="callout insight">
<div class="callout-title">๐ Traditional DB vs Vector DB</div>
<strong>Traditional (SQL):</strong> "Find documents with the word 'dog'" โ Exact keyword match<br>
<strong>Vector DB:</strong> "Find documents about pets" โ Returns articles about dogs, cats, hamsters โ because their <em>meaning</em> is similar even if they don't share exact words.
</div>
<h3>How It Works</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Embed:</strong> Convert text/images into vectors using an embedding model (e.g., OpenAI ada-002, sentence-transformers)</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Index:</strong> Store vectors in a specialized data structure (HNSW, IVF, product quantization) for fast search</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Query:</strong> Convert query to a vector, find the K nearest neighbors using cosine similarity or L2 distance</div></div>
<h3>Popular Vector Databases</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Database</th>
<th style="padding: 10px; color: var(--cyan);">Type</th>
<th style="padding: 10px; color: var(--cyan);">Best For</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">ChromaDB</td><td>In-memory, local</td><td>Prototyping, small datasets</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">FAISS (Meta)</td><td>Library</td><td>Billion-scale search, GPU-accelerated</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Pinecone</td><td>Cloud service</td><td>Production SaaS, managed infrastructure</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Weaviate</td><td>Open-source</td><td>Multi-modal search, GraphQL API</td></tr>
<tr><td style="padding: 8px;">Milvus</td><td>Open-source</td><td>Enterprise, distributed scaling</td></tr>
</table>
`,
concepts: `
<h3>Core Concepts</h3>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Embeddings:</strong> Dense numerical representations of data (text, images, etc.) produced by neural networks. Example: OpenAI's text-embedding-3-small produces 1536-dim vectors.</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Similarity Metrics:</strong> Cosine similarity (angle), Euclidean distance (L2), Dot product (magnitude-aware). Cosine is most common for text embeddings.</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>ANN Algorithms:</strong> Approximate Nearest Neighbor โ trade small accuracy loss for massive speed gains (exact search is O(n), ANN is O(log n)).</div>
</div>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>HNSW (Hierarchical Navigable Small World):</strong> Graph-based index. Builds a multi-layer graph where higher layers have fewer, long-range connections. Most popular algorithm.</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>IVF (Inverted File Index):</strong> Clusters vectors first (like k-means), then only searches relevant clusters. Fast but requires training step.</div>
</div>
<div class="list-item">
<div class="list-num">06</div>
<div><strong>Product Quantization (PQ):</strong> Compresses vectors by splitting them into sub-vectors and quantizing each. Reduces memory by 10-100x.</div>
</div>
<h3>Embedding Pipeline</h3>
<div class="formul a">
1. Raw Data โ Embedding Model โ Vector (e.g., [0.12, -0.34, 0.56, ...])<br>
2. Store vector + metadata in Vector DB<br>
3. Query: Convert query โ vector โ Find nearest neighbors<br>
4. Return top-k most similar results with scores
</div>
<div class="callout warning">
<div class="callout-title">โ ๏ธ Common Pitfalls</div>
โข Using wrong distance metric (cosine for normalized, L2 for raw)<br>
โข Not chunking documents properly (too large = diluted embedding)<br>
โข Mixing embedding models (query and docs must use same model)<br>
โข Ignoring metadata filtering (combine vector search + filters)
</div>
`,
math: `
<h3>๐ Paper & Pain: Vector Similarity Mathematics</h3>
<h4>Cosine Similarity</h4>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>cos(A, B) = (A ยท B) / (||A|| ร ||B||)</strong>
</div>
<div class="callout insight">
<div class="callout-title">๐ Manual Calculation</div>
<strong>A = [1, 2, 3], B = [4, 5, 6]</strong><br><br>
<strong>Step 1 โ Dot Product:</strong><br>
A ยท B = (1ร4) + (2ร5) + (3ร6) = 4 + 10 + 18 = <strong>32</strong><br><br>
<strong>Step 2 โ Magnitudes:</strong><br>
||A|| = โ(1ยฒ + 2ยฒ + 3ยฒ) = โ14 โ 3.742<br>
||B|| = โ(4ยฒ + 5ยฒ + 6ยฒ) = โ77 โ 8.775<br><br>
<strong>Step 3 โ Cosine Similarity:</strong><br>
cos(A, B) = 32 / (3.742 ร 8.775) = 32 / 32.833 โ <strong>0.9746</strong><br><br>
<strong>Interpretation:</strong> Very high similarity (close to 1.0 = identical direction)
</div>
<h4>Euclidean Distance (L2)</h4>
<div class="formula">
d(A, B) = โ(ฮฃ(aแตข - bแตข)ยฒ)<br><br>
For A = [1, 2, 3], B = [4, 5, 6]:<br>
d = โ((4-1)ยฒ + (5-2)ยฒ + (6-3)ยฒ) = โ(9 + 9 + 9) = โ27 โ <strong>5.196</strong>
</div>
<h4>HNSW Complexity</h4>
<div class="formula">
Build time: O(N ร log(N))<br>
Query time: O(log(N)) โ logarithmic!<br>
Memory: O(N ร M) where M = max connections per node<br><br>
Compare to brute force: O(N) per query<br>
For N = 1 billion vectors: HNSW is ~30 orders of magnitude faster
</div>
`,
applications: `
<h3>Real-World Applications</h3>
<div class="info-box">
<div class="box-title">๐ Semantic Search</div>
<div class="box-content">
<strong>Google, Bing, Notion AI:</strong> Search by meaning, not keywords<br>
<strong>Example:</strong> "How to fix my car" matches "automobile repair guide" even with zero keyword overlap
</div>
</div>
<div class="info-box">
<div class="box-title">๐ค RAG (Retrieval-Augmented Generation)</div>
<div class="box-content">
Store knowledge base as vectors โ retrieve relevant context โ feed to LLM for grounded answers. Used by ChatGPT with browsing, Perplexity AI, and enterprise AI assistants.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ Recommendation Systems</div>
<div class="box-content">
<strong>Spotify (music):</strong> Embed songs as vectors, recommend nearest neighbors<br>
<strong>Netflix (movies):</strong> User + content embeddings for personalization
</div>
</div>
<div class="info-box">
<div class="box-title">๐ผ๏ธ Image & Multi-Modal Search</div>
<div class="box-content">
<strong>CLIP embeddings:</strong> Search images with text queries and vice versa<br>
<strong>Google Lens:</strong> Find products by uploading a photo
</div>
</div>
`,
code: `
<h3>Similarity Metrics</h3>
<div class="formula" style="text-align:center;">
<strong>Cosine Similarity:</strong> $$\\cos(\\theta) = \\frac{A \\cdot B}{\\|A\\| \\|B\\|}$$
<strong>L2 (Euclidean):</strong> $$d = \\sqrt{\\sum(a_i - b_i)^2}$$
</div>
<span class="code-title">๐ chromadb_example.py</span><div class="code-block"><span class="keyword">import</span> chromadb
client = chromadb.Client()
collection = client.create_collection(<span class="string">"dl_docs"</span>)
<span class="comment"># Add documents</span>
collection.add(
documents=[<span class="string">"Neural networks use backpropagation"</span>, <span class="string">"CNNs detect visual patterns"</span>],
ids=[<span class="string">"doc1"</span>, <span class="string">"doc2"</span>],
)
<span class="comment"># Query</span>
results = collection.query(query_texts=[<span class="string">"How do neural nets learn?"</span>], n_results=<span class="number">1</span>)
<span class="keyword">print</span>(results[<span class="string">"documents"</span>]) <span class="comment"># [["Neural networks use backpropagation"]]</span></div>
`
},
"rag": {
overview: `
<p>RAG (Retrieval-Augmented Generation) is the <strong>most practical pattern for building AI applications</strong> today. Instead of relying solely on an LLM's training data (which is frozen at training time), RAG retrieves relevant documents from your own knowledge base and includes them in the prompt.</p>
<div class="callout insight">
<div class="callout-title">๐ก Why RAG Matters</div>
LLMs have two fundamental problems: (1) they hallucinate โ confidently generating false information, and (2) their knowledge has a cutoff date. RAG solves both by grounding the LLM's responses in <strong>actual, verifiable documents</strong> from your data.
</div>
<h3>The RAG Pipeline</h3>
<div class="list-item"><div class="list-num">1</div><div><strong>Ingestion (Offline):</strong> Load documents โ Split into chunks (500-1000 tokens) โ Embed each chunk โ Store in vector DB</div></div>
<div class="list-item"><div class="list-num">2</div><div><strong>Retrieval (Online):</strong> User asks a question โ Embed the question โ Find K most similar chunks in vector DB</div></div>
<div class="list-item"><div class="list-num">3</div><div><strong>Generation:</strong> Stuff retrieved chunks into the prompt as context โ LLM generates an answer grounded in the context</div></div>
<h3>RAG vs. Fine-Tuning</h3>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Aspect</th>
<th style="padding: 10px; color: var(--cyan);">RAG</th>
<th style="padding: 10px; color: var(--cyan);">Fine-Tuning</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Knowledge updates</td><td>Instant (add new docs)</td><td>Requires retraining</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Hallucination</td><td>Reduced (grounded in docs)</td><td>Can still hallucinate</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Cost</td><td>Low (no training needed)</td><td>High (GPU compute)</td></tr>
<tr><td style="padding: 8px;">Best for</td><td>QA over docs, support bots</td><td>Teaching new skills/style</td></tr>
</table>
`,
concepts: `
<h3>RAG Architecture Deep Dive</h3>
<h4>1. Document Processing</h4>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Loading:</strong> Parse PDFs, HTML, Markdown, Databases, APIs using document loaders (LangChain, LlamaIndex)</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Chunking Strategies:</strong><br>
โข <strong>Fixed-size:</strong> Split every N tokens (simplest, often good enough)<br>
โข <strong>Recursive:</strong> Split by paragraphs โ sentences โ words (preserves structure)<br>
โข <strong>Semantic:</strong> Use embeddings to detect topic boundaries<br>
โข <strong>Overlap:</strong> Add 10-20% overlap between chunks to preserve context
</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Embedding:</strong> Convert chunks to vectors using models like OpenAI text-embedding-3-small (1536-dim), Cohere embed-v3, or open-source models like BGE-large</div>
</div>
<h4>2. Retrieval Strategies</h4>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>Dense Retrieval:</strong> Cosine similarity on embeddings (semantic search)</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Sparse Retrieval:</strong> BM25/TF-IDF keyword matching (exact term matching)</div>
</div>
<div class="list-item">
<div class="list-num">06</div>
<div><strong>Hybrid Search:</strong> Combine dense + sparse with Reciprocal Rank Fusion (RRF) โ usually best results</div>
</div>
<h4>3. Advanced RAG Patterns</h4>
<div class="list-item">
<div class="list-num">07</div>
<div><strong>Multi-Query RAG:</strong> LLM generates multiple query variations โ retrieve for each โ merge results</div>
</div>
<div class="list-item">
<div class="list-num">08</div>
<div><strong>Re-ranking:</strong> Use a cross-encoder (e.g., Cohere Rerank) to re-score retrieved chunks for higher relevance</div>
</div>
<div class="list-item">
<div class="list-num">09</div>
<div><strong>Query Expansion:</strong> HyDE โ generate a hypothetical answer, embed that, then retrieve</div>
</div>
<div class="list-item">
<div class="list-num">10</div>
<div><strong>Agentic RAG:</strong> LLM decides when and what to retrieve; can self-reflect and re-retrieve if initial results are poor</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ Production Checklist</div>
โ
Chunk size 256-512 tokens with 10-20% overlap<br>
โ
Use hybrid search (dense + BM25)<br>
โ
Add re-ranker for precision<br>
โ
Include metadata filtering (date, source, category)<br>
โ
Monitor retrieval quality with evaluation metrics
</div>
`,
math: `
<h3>๐ RAG Evaluation Metrics</h3>
<h4>Retrieval Quality</h4>
<div class="formula" style="font-size: 1.1rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>Recall@k = |Relevant โฉ Retrieved@k| / |Relevant|</strong><br><br>
<strong>Precision@k = |Relevant โฉ Retrieved@k| / k</strong><br><br>
<strong>MRR = 1/|Q| ร ฮฃ(1 / rank_i)</strong>
</div>
<div class="callout insight">
<div class="callout-title">๐ Paper & Pain: MRR Calculation</div>
<strong>3 queries, first relevant result at ranks 1, 3, 2:</strong><br><br>
MRR = (1/3) ร (1/1 + 1/3 + 1/2)<br>
MRR = (1/3) ร (1.0 + 0.333 + 0.5)<br>
MRR = (1/3) ร 1.833 = <strong>0.611</strong><br><br>
<strong>Interpretation:</strong> On average, the first relevant result appears around position 1.6
</div>
<h4>Reciprocal Rank Fusion (Hybrid Search)</h4>
<div class="formula">
<strong>RRF(d) = ฮฃ 1 / (k + rank_i(d))</strong><br><br>
Where k = 60 (constant), rank_i = rank from retriever i<br><br>
Example: Document D appears at rank 2 in dense, rank 5 in sparse:<br>
RRF(D) = 1/(60+2) + 1/(60+5) = 0.0161 + 0.0154 = <strong>0.0315</strong>
</div>
<h4>Chunking Size Impact</h4>
<div class="formula">
Optimal chunk size depends on:<br>
โข Embedding model context window (usually 512 tokens max)<br>
โข Query specificity (specific โ smaller chunks)<br>
โข Document structure (code โ function-level, prose โ paragraph-level)<br><br>
Rule of thumb: <strong>chunk_size โ 2-3ร expected query length</strong>
</div>
`,
applications: `
<h3>RAG in Production</h3>
<div class="info-box">
<div class="box-title">๐ Perplexity AI</div>
<div class="box-content">
RAG-powered search engine: retrieves web pages โ feeds to LLM โ generates cited answers. Processes 100M+ queries/month.
</div>
</div>
<div class="info-box">
<div class="box-title">๐ผ Enterprise Knowledge Bases</div>
<div class="box-content">
<strong>Notion AI, Glean, Guru:</strong> Index company docs โ answer employee questions with source citations<br>
<strong>Legal AI:</strong> Retrieve relevant case law for legal research (Harvey AI)
</div>
</div>
<div class="info-box">
<div class="box-title">๐ป Code Assistants</div>
<div class="box-content">
<strong>GitHub Copilot, Cursor:</strong> Retrieve relevant code from the codebase to inform suggestions<br>
<strong>Documentation Q&A:</strong> Answer questions about SDKs and libraries
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฅ Healthcare</div>
<div class="box-content">
Retrieve medical literature โ generate evidence-based clinical summaries (Google Med-PaLM 2)
</div>
</div>
<div class="callout tip">
<div class="callout-title">๐ ๏ธ Popular Frameworks</div>
โข <strong>LangChain:</strong> Most popular, modular pipeline builder<br>
โข <strong>LlamaIndex:</strong> Data-focused, great for document processing<br>
โข <strong>Haystack:</strong> Production-ready, NLP-first<br>
โข <strong>Semantic Kernel:</strong> Microsoft's enterprise RAG framework
</div>
`,
code: `<span class="code-title">๐ rag_langchain.py</span><div class="code-block"><span class="keyword">from</span> langchain.document_loaders <span class="keyword">import</span> PyPDFLoader
<span class="keyword">from</span> langchain.text_splitter <span class="keyword">import</span> RecursiveCharacterTextSplitter
<span class="keyword">from</span> langchain.embeddings <span class="keyword">import</span> OpenAIEmbeddings
<span class="keyword">from</span> langchain.vectorstores <span class="keyword">import</span> Chroma
<span class="keyword">from</span> langchain.chains <span class="keyword">import</span> RetrievalQA
<span class="keyword">from</span> langchain.llms <span class="keyword">import</span> OpenAI
<span class="comment"># 1. Load & chunk documents</span>
loader = PyPDFLoader(<span class="string">"deep_learning_book.pdf"</span>)
docs = loader.load()
splitter = RecursiveCharacterTextSplitter(chunk_size=<span class="number">1000</span>, chunk_overlap=<span class="number">200</span>)
chunks = splitter.split_documents(docs)
<span class="comment"># 2. Create vector store</span>
embeddings = OpenAIEmbeddings()
vectordb = Chroma.from_documents(chunks, embeddings, persist_directory=<span class="string">"./chroma_db"</span>)
<span class="comment"># 3. Query with RAG</span>
qa_chain = RetrievalQA.from_chain_type(
llm=OpenAI(temperature=<span class="number">0</span>),
retriever=vectordb.as_retriever(search_kwargs={<span class="string">"k"</span>: <span class="number">3</span>}),
)
answer = qa_chain.run(<span class="string">"What is backpropagation?"</span>)
<span class="keyword">print</span>(answer)</div>
`
},
"advanced-llm": {
overview: `
<p>Fine-tuning a 7B+ parameter model requires enormous GPU memory (28GB+ for fp32). <strong>LoRA (Low-Rank Adaptation)</strong> makes this practical by training only a tiny fraction of parameters โ typically <strong>0.1% or less</strong> โ while achieving similar performance to full fine-tuning.</p>
<div class="callout insight">
<div class="callout-title">๐ฏ The Key Insight</div>
Research showed that pretrained weight matrices have <strong>low intrinsic dimensionality</strong> โ meaning changes during fine-tuning can be represented by low-rank matrices. Instead of updating the full dรd weight matrix W, we add a small update BA where B is dรr and A is rรd, with rank r << d.
</div>
<h3>LoRA Math</h3>
<div class="formula" style="text-align:center;">
$$W' = W + \\Delta W = W + BA$$
$$B \\in \\mathbb{R}^{d \\times r}, \\; A \\in \\mathbb{R}^{r \\times d}, \\; r \\ll d$$
</div>
<p><strong>Example:</strong> For a 4096ร4096 weight matrix (16.7M params), LoRA with r=16 adds only 2ร4096ร16 = 131K params โ a <strong>128x reduction</strong>.</p>
<h3>Quantization + LoRA = QLoRA</h3>
<p>QLoRA takes this further by keeping the base model in <strong>4-bit quantized</strong> format and training LoRA adapters in fp16. This means fine-tuning a 65B model on a <strong>single 48GB GPU</strong>.</p>
<table style="width:100%; border-collapse: collapse; margin: 15px 0;">
<tr style="border-bottom: 2px solid var(--cyan);">
<th style="padding: 10px; color: var(--cyan);">Method</th>
<th style="padding: 10px; color: var(--cyan);">7B Model Memory</th>
<th style="padding: 10px; color: var(--cyan);">Trainable Params</th>
</tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Full Fine-tune (fp32)</td><td>28 GB</td><td>7B (100%)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">Full Fine-tune (fp16)</td><td>14 GB</td><td>7B (100%)</td></tr>
<tr style="border-bottom: 1px solid rgba(255,255,255,0.1);"><td style="padding: 8px;">LoRA (fp16)</td><td>14 GB</td><td>~4M (0.06%)</td></tr>
<tr><td style="padding: 8px;">QLoRA (4-bit)</td><td>~4 GB</td><td>~4M (0.06%)</td></tr>
</table>
`,
concepts: `
<h3>LoRA (Low-Rank Adaptation)</h3>
<p>Instead of updating all parameters, LoRA freezes the original weights and injects small trainable matrices into each layer.</p>
<div class="list-item">
<div class="list-num">01</div>
<div><strong>Core Idea:</strong> Weight update ฮW can be decomposed as a low-rank matrix: ฮW = B ร A, where B โ โ^(dรr) and A โ โ^(rรd), with r โช d</div>
</div>
<div class="list-item">
<div class="list-num">02</div>
<div><strong>Memory Savings:</strong> For a 7B model, full fine-tuning needs ~28GB VRAM. LoRA (r=16) needs ~6GB โ a 4.7x reduction.</div>
</div>
<div class="list-item">
<div class="list-num">03</div>
<div><strong>Rank (r):</strong> Typical values: r=8, 16, 32, 64. Higher rank = more expressiveness but more parameters.</div>
</div>
<h3>QLoRA (Quantized LoRA)</h3>
<div class="list-item">
<div class="list-num">04</div>
<div><strong>4-bit NormalFloat (NF4):</strong> Quantize frozen weights to 4-bit. Train LoRA adapters in fp16/bf16. Enables fine-tuning 65B models on a single 48GB GPU.</div>
</div>
<div class="list-item">
<div class="list-num">05</div>
<div><strong>Double Quantization:</strong> Quantize the quantization constants themselves โ saves additional memory.</div>
</div>
<h3>Quantization for Deployment</h3>
<div class="list-item">
<div class="list-num">06</div>
<div><strong>GGUF (llama.cpp):</strong> CPU-optimized format. Run Llama-2 70B on a MacBook with Q4_K_M quantization.</div>
</div>
<div class="list-item">
<div class="list-num">07</div>
<div><strong>GPTQ:</strong> GPU-optimized post-training quantization. 3-4 bit with minimal quality loss.</div>
</div>
<div class="list-item">
<div class="list-num">08</div>
<div><strong>AWQ (Activation-aware):</strong> Preserves important weight channels. State-of-the-art quality at 4-bit.</div>
</div>
<h3>RLHF & DPO</h3>
<div class="list-item">
<div class="list-num">09</div>
<div><strong>RLHF:</strong> Train a reward model from human preferences, then use PPO to optimize the LLM's policy. Used by ChatGPT, Claude.</div>
</div>
<div class="list-item">
<div class="list-num">10</div>
<div><strong>DPO (Direct Preference Optimization):</strong> Skip the reward model entirely โ optimize directly from preference pairs. Simpler, lower compute, often comparable results.</div>
</div>
<div class="callout insight">
<div class="callout-title">๐ PEFT (Parameter-Efficient Fine-Tuning)</div>
PEFT is the umbrella term for all methods that train <1% of parameters:<br>
โข <strong>LoRA / QLoRA:</strong> Low-rank weight decomposition<br>
โข <strong>Prefix Tuning:</strong> Trainable prefix tokens<br>
โข <strong>Adapters:</strong> Small bottleneck layers inserted into frozen model<br>
โข <strong>IA3:</strong> Learned rescaling vectors (even fewer params than LoRA)
</div>
`,
math: `
<h3>๐ Paper & Pain: LoRA Mathematics</h3>
<h4>Low-Rank Decomposition</h4>
<div class="formula" style="font-size: 1.2rem; text-align: center; margin: 20px 0; background: rgba(0, 212, 255, 0.08); padding: 25px; border-radius: 8px;">
<strong>W' = W + ฮW = W + B ร A</strong><br>
<small>where W โ โ^(dรd), B โ โ^(dรr), A โ โ^(rรd), r โช d</small>
</div>
<div class="callout insight">
<div class="callout-title">๐ Parameter Count Comparison</div>
<strong>Full fine-tuning of one attention layer (d=4096):</strong><br>
Parameters = d ร d = 4096 ร 4096 = <strong>16,777,216</strong><br><br>
<strong>LoRA (r=16):</strong><br>
Parameters = d ร r + r ร d = 4096 ร 16 + 16 ร 4096 = <strong>131,072</strong><br><br>
<strong>Reduction: 128x fewer parameters!</strong> (0.78% of original)<br><br>
For a full 7B model (all attention layers):<br>
Full: ~7B params to train (~28GB VRAM)<br>
LoRA (r=16): ~4.2M params to train (~1.5GB for adapters)
</div>
<h4>Quantization Math</h4>
<div class="formula">
<strong>Linear Quantization (INT8):</strong><br>
q = round(w / scale + zero_point)<br>
w_approx = (q - zero_point) ร scale<br><br>
<strong>Example:</strong><br>
Weight w = 0.73, scale = 0.01, zero_point = 128<br>
q = round(0.73 / 0.01 + 128) = round(201) = 201<br>
w_approx = (201 - 128) ร 0.01 = 0.73 โ
</div>
<h4>Memory Savings</h4>
<div class="formula">
<strong>Model Size = Parameters ร Bytes per Parameter</strong><br><br>
7B model at fp32: 7B ร 4 bytes = 28 GB<br>
7B model at fp16: 7B ร 2 bytes = 14 GB<br>
7B model at INT8: 7B ร 1 byte = 7 GB<br>
7B model at INT4: 7B ร 0.5 bytes = 3.5 GB<br><br>
<strong>4-bit quantization = 8x memory reduction!</strong>
</div>
<h4>DPO Loss Function</h4>
<div class="formula">
L_DPO = -E[log ฯ(ฮฒ(log ฯ_ฮธ(y_w|x) - log ฯ_ref(y_w|x) - log ฯ_ฮธ(y_l|x) + log ฯ_ref(y_l|x)))]<br><br>
Where:<br>
โข y_w = preferred response, y_l = rejected response<br>
โข ฯ_ฮธ = policy model, ฯ_ref = reference model<br>
โข ฮฒ = temperature controlling deviation from reference
</div>
`,
applications: `
<h3>Production Deployment</h3>
<div class="info-box">
<div class="box-title">๐ Local LLM Deployment</div>
<div class="box-content">
<strong>Ollama:</strong> One-command local LLM deployment (ollama run llama3)<br>
<strong>llama.cpp:</strong> CPU inference with GGUF quantized models<br>
<strong>vLLM:</strong> High-throughput GPU serving with PagedAttention
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฏ Domain-Specific Fine-Tuning</div>
<div class="box-content">
<strong>Medical:</strong> BioMistral, Med-PaLM (clinical notes, diagnoses)<br>
<strong>Legal:</strong> SaulLM (contract analysis, case law)<br>
<strong>Code:</strong> CodeLlama, StarCoder (code generation, completion)
</div>
</div>
<div class="info-box">
<div class="box-title">๐ฐ Cost Optimization</div>
<div class="box-content">
<strong>Distillation:</strong> Train a small model to mimic a large one (GPT-4 โ Phi-3)<br>
<strong>Speculative Decoding:</strong> Small model drafts, large model verifies (2-3x speedup)<br>
<strong>Mixture of Experts:</strong> Only activate 2 of 8 expert networks per token (Mixtral)
</div>
</div>
<div class="info-box">
<div class="box-title">๐ ๏ธ Key Tools</div>
<div class="box-content">
<strong>Hugging Face PEFT:</strong> Official LoRA/QLoRA library<br>
<strong>Unsloth:</strong> 2x faster fine-tuning with 60% less memory<br>
<strong>Axolotl:</strong> Config-driven fine-tuning (no code needed)<br>
<strong>TRL:</strong> Transformer Reinforcement Learning (RLHF/DPO)
</div>
</div>
`,
code: `
<h3>LoRA Math</h3>
<div class="formula" style="text-align:center;">
$$W' = W + BA, \\quad B \\in \\mathbb{R}^{d \\times r}, \\; A \\in \\mathbb{R}^{r \\times d}, \\; r \\ll d$$
Trainable params: $2 \\times d \\times r$ instead of $d^2$ (e.g., <strong>128x reduction</strong>)
</div>
<span class="code-title">๐ lora_finetuning.py</span><div class="code-block"><span class="keyword">from</span> peft <span class="keyword">import</span> LoraConfig, get_peft_model
<span class="keyword">from</span> transformers <span class="keyword">import</span> AutoModelForCausalLM, AutoTokenizer
<span class="comment"># Load base model</span>
model = AutoModelForCausalLM.from_pretrained(<span class="string">"meta-llama/Llama-2-7b-hf"</span>)
<span class="comment"># Apply LoRA</span>
lora_config = LoraConfig(
r=<span class="number">16</span>, <span class="comment"># rank (lower = fewer params)</span>
lora_alpha=<span class="number">32</span>, <span class="comment"># scaling factor</span>
target_modules=[<span class="string">"q_proj"</span>, <span class="string">"v_proj"</span>], <span class="comment"># which layers to adapt</span>
lora_dropout=<span class="number">0.05</span>,
task_type=<span class="string">"CAUSAL_LM"</span>,
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
<span class="comment"># Output: trainable: 4,194,304 || all: 6,742,609,920 || 0.06%</span></div>
`
}
};
function createModuleHTML(module) {
const content = MODULE_CONTENT[module.id] || {};
return `
<div class="module" id="${module.id}-module">
<button class="btn-back" onclick="switchTo('dashboard')">โ Back to Dashboard</button>
<header>
<h1>${module.icon} ${module.title}</h1>
<p class="subtitle">${module.description}</p>
</header>
<div class="tabs">
<button class="tab-btn active" onclick="switchTab(event, '${module.id}-overview')">Overview</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-concepts')">Key Concepts</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-visualization')">๐ Visualization</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-math')">Math</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-applications')">Applications</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-code')">๐ Python Code</button>
<button class="tab-btn" onclick="switchTab(event, '${module.id}-summary')">Summary</button>
</div>
<div id="${module.id}-overview" class="tab active">
<div class="section">
<h2>๐ Overview</h2>
${content.overview || `
<p>Complete coverage of ${module.title.toLowerCase()}. Learn the fundamentals, mathematics, real-world applications, and implementation details.</p>
<div class="info-box">
<div class="box-title">Learning Objectives</div>
<div class="box-content">
โ Understand core concepts and theory<br>
โ Master mathematical foundations<br>
โ Learn practical applications<br>
โ Implement and experiment
</div>
</div>
`}
</div>
</div>
<div id="${module.id}-concepts" class="tab">
<div class="section">
<h2>๐ฏ Key Concepts</h2>
${content.concepts || `
<p>Fundamental concepts and building blocks for ${module.title.toLowerCase()}.</p>
<div class="callout insight">
<div class="callout-title">๐ก Main Ideas</div>
This section covers the core ideas you need to understand before diving into mathematics.
</div>
`}
</div>
</div>
<div id="${module.id}-visualization" class="tab">
<div class="section">
<h2>๐ Interactive Visualization</h2>
<p>Visual representation to help understand ${module.title.toLowerCase()} concepts intuitively.</p>
<div id="${module.id}-viz" class="viz-container">
<canvas id="${module.id}-canvas" width="800" height="400" style="border: 1px solid rgba(0, 212, 255, 0.3); border-radius: 8px; background: rgba(0, 212, 255, 0.02);"></canvas>
</div>
<div class="viz-controls">
<button onclick="drawVisualization('${module.id}')" class="btn-viz">๐ Refresh Visualization</button>
<button onclick="toggleVizAnimation('${module.id}')" class="btn-viz">โถ๏ธ Animate</button>
<button onclick="downloadViz('${module.id}')" class="btn-viz">โฌ๏ธ Save Image</button>
</div>
</div>
</div>
<div id="${module.id}-math" class="tab">
<div class="section">
<h2>๐ Mathematical Foundation</h2>
${content.math || `
<p>Rigorous mathematical treatment of ${module.title.toLowerCase()}.</p>
<div class="formula">
Mathematical formulas and derivations go here
</div>
`}
</div>
</div>
<div id="${module.id}-applications" class="tab">
<div class="section">
<h2>๐ Real-World Applications</h2>
${content.applications || `
<p>How ${module.title.toLowerCase()} is used in practice across different industries.</p>
<div class="info-box">
<div class="box-title">Use Cases</div>
<div class="box-content">
Common applications and practical examples
</div>
</div>
`}
</div>
</div>
<div id="${module.id}-code" class="tab">
<div class="section">
<h2>๐ Python Implementation</h2>
${content.code || `
<p>Production-ready Python code for ${module.title.toLowerCase()} using PyTorch and TensorFlow.</p>
<div class="callout tip">
<div class="callout-title">๐ก Coming Soon</div>
Python code examples for this module are being prepared.
</div>
`}
</div>
</div>
<div id="${module.id}-summary" class="tab">
<div class="section">
<h2>โ
Summary</h2>
<div class="info-box">
<div class="box-title">Key Takeaways</div>
<div class="box-content">
โ Essential concepts covered<br>
โ Mathematical foundations understood<br>
โ Real-world applications identified<br>
โ Ready for implementation
</div>
</div>
</div>
</div>
</div>
`;
}
function initDashboard() {
const grid = document.getElementById("modulesGrid");
const container = document.getElementById("modulesContainer");
modules.forEach((module, index) => {
const card = document.createElement("div");
// Add staggered animation class
const staggerClass = `stagger stagger-${(index % 8) + 1}`;
card.className = `card hover-glow ${staggerClass}`;
card.style.borderColor = module.color;
card.onclick = () => switchTo(module.id + "-module");
card.innerHTML = `
<div class="card-icon">${module.icon}</div>
<h3>${module.title}</h3>
<p>${module.description}</p>
<span class="category-label">${module.category}</span>
`;
grid.appendChild(card);
const moduleHTML = createModuleHTML(module);
container.innerHTML += moduleHTML;
});
}
function switchTo(target) {
document.querySelectorAll('.dashboard, .module').forEach(el => {
el.classList.remove('active');
});
const elem = document.getElementById(target);
if (elem) elem.classList.add('active');
}
function switchTab(e, tabId) {
const module = e.target.closest('.module');
if (!module) return;
module.querySelectorAll('.tab').forEach(t => t.classList.remove('active'));
module.querySelectorAll('.tab-btn').forEach(b => b.classList.remove('active'));
const tab = document.getElementById(tabId);
if (tab) tab.classList.add('active');
e.target.classList.add('active');
// Re-render MathJax in the newly visible tab
if (window.MathJax && window.MathJax.typesetPromise) {
window.MathJax.typesetPromise([tab]).catch(function (err) { console.log('MathJax typeset:', err); });
}
// Trigger visualization when tabs are clicked
setTimeout(() => {
const moduleId = tabId.split('-')[0];
if (tabId.includes('-concepts')) {
drawConceptsVisualization(moduleId);
} else if (tabId.includes('-visualization')) {
drawConceptsVisualization(moduleId);
} else if (tabId.includes('-math')) {
drawMathVisualization(moduleId);
} else if (tabId.includes('-applications')) {
drawApplicationVisualization(moduleId);
}
}, 150);
}
// Visualization Functions - Concepts Tab
function drawConceptsVisualization(moduleId) {
const canvas = document.getElementById(moduleId + '-canvas');
if (!canvas) return;
const ctx = canvas.getContext('2d');
ctx.clearRect(0, 0, canvas.width, canvas.height);
ctx.fillStyle = '#0f1419';
ctx.fillRect(0, 0, canvas.width, canvas.height);
const vizMap = {
'nn-basics': drawNeuronAnimation,
'perceptron': drawDecisionBoundary,
'mlp': drawNetworkGraph,
'activation': drawActivationFunctions,
'weight-init': drawWeightDistribution,
'loss': drawLossLandscape,
'optimizers': drawConvergencePaths,
'backprop': drawGradientFlow,
'regularization': drawOverfitComparison,
'batch-norm': drawBatchNormalization,
'cv-intro': drawImageMatrix,
'conv-layer': drawConvolutionAnimation,
'pooling': drawPoolingDemo,
'cnn-basics': drawCNNArchitecture,
'viz-filters': drawLearnedFilters,
'lenet': drawLeNetArchitecture,
'alexnet': drawAlexNetArchitecture,
'vgg': drawVGGArchitecture,
'resnet': drawResNetArchitecture,
'inception': drawInceptionModule,
'mobilenet': drawMobileNetArchitecture,
'transfer-learning': drawTransferLearning,
'localization': drawBoundingBoxes,
'rcnn': drawRCNNPipeline,
'yolo': drawYOLOGrid,
'ssd': drawSSDDetector,
'semantic-seg': drawSemanticSegmentation,
'instance-seg': drawInstanceSegmentation,
'face-recog': drawFaceEmbeddings,
'autoencoders': drawAutoencoderArchitecture,
'gans': drawGANsGame,
'diffusion': drawDiffusionProcess,
'rnn': drawRNNUnrolled,
'transformers': drawAttentionMatrix,
'bert': drawBERTProcess,
'gpt': drawGPTGeneration,
'vit': drawVisionTransformer,
'gnn': drawGraphNetwork,
'seq2seq': drawSeq2SeqAttention,
'research-papers': drawDefaultVisualization,
'vector-db': drawVectorSpace,
'rag': drawRAGPipeline,
'advanced-llm': drawLoRADiagram
};
if (vizMap[moduleId]) {
vizMap[moduleId](ctx, canvas);
} else {
drawDefaultVisualization(ctx, canvas);
}
}
// Default Visualization
function drawDefaultVisualization(ctx, canvas) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
ctx.fillStyle = 'rgba(0, 212, 255, 0.2)';
ctx.fillRect(centerX - 120, centerY - 60, 240, 120);
ctx.strokeStyle = '#00d4ff';
ctx.lineWidth = 2;
ctx.strokeRect(centerX - 120, centerY - 60, 240, 120);
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 18px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Interactive Visualization', centerX, centerY - 20);
ctx.font = '13px Arial';
ctx.fillText('Custom visualization for this topic', centerX, centerY + 20);
ctx.font = '11px Arial';
ctx.fillStyle = '#00ff88';
ctx.fillText('Click Refresh to render', centerX, centerY + 45);
}
// Default Math Visualization
function drawDefaultMathVisualization(ctx, canvas) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
ctx.fillStyle = 'rgba(255, 107, 53, 0.2)';
ctx.fillRect(centerX - 120, centerY - 60, 240, 120);
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 2;
ctx.strokeRect(centerX - 120, centerY - 60, 240, 120);
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 18px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Mathematical Formulas', centerX, centerY - 20);
ctx.font = '13px Arial';
ctx.fillText('Visual equation derivations', centerX, centerY + 20);
ctx.font = '11px Arial';
ctx.fillStyle = '#00ff88';
ctx.fillText('Click Visualize to render', centerX, centerY + 45);
}
// Default Application Visualization
function drawDefaultApplicationVisualization(ctx, canvas) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
ctx.fillStyle = 'rgba(0, 255, 136, 0.2)';
ctx.fillRect(centerX - 120, centerY - 60, 240, 120);
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 2;
ctx.strokeRect(centerX - 120, centerY - 60, 240, 120);
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 18px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Real-World Applications', centerX, centerY - 20);
ctx.font = '13px Arial';
ctx.fillText('Practical use cases and examples', centerX, centerY + 20);
ctx.font = '11px Arial';
ctx.fillStyle = '#ffa500';
ctx.fillText('Click Show Applications to render', centerX, centerY + 45);
}
// Activation Functions Visualization
function drawActivationFunctions(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
const centerX = width / 2;
const centerY = height / 2;
const scale = 40;
// Draw grid
ctx.strokeStyle = 'rgba(0, 212, 255, 0.1)';
ctx.lineWidth = 1;
for (let i = -5; i <= 5; i += 1) {
const x = centerX + i * scale;
ctx.beginPath();
ctx.moveTo(x, centerY - 5 * scale);
ctx.lineTo(x, centerY + 5 * scale);
ctx.stroke();
}
// Draw axes
ctx.strokeStyle = '#00d4ff';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(centerX - 6 * scale, centerY);
ctx.lineTo(centerX + 6 * scale, centerY);
ctx.stroke();
ctx.beginPath();
ctx.moveTo(centerX, centerY - 6 * scale);
ctx.lineTo(centerX, centerY + 6 * scale);
ctx.stroke();
// Draw activation functions
const functions = [
{ name: 'ReLU', color: '#ff6b35', fn: x => Math.max(0, x) },
{ name: 'Sigmoid', color: '#00ff88', fn: x => 1 / (1 + Math.exp(-x)) },
{ name: 'Tanh', color: '#ffa500', fn: x => Math.tanh(x) }
];
functions.forEach(func => {
ctx.strokeStyle = func.color;
ctx.lineWidth = 2;
ctx.beginPath();
for (let x = -5; x <= 5; x += 0.1) {
const y = func.fn(x);
const canvasX = centerX + x * scale;
const canvasY = centerY - y * scale;
if (x === -5) ctx.moveTo(canvasX, canvasY);
else ctx.lineTo(canvasX, canvasY);
}
ctx.stroke();
});
// Legend
ctx.font = 'bold 12px Arial';
functions.forEach((func, i) => {
ctx.fillStyle = func.color;
ctx.fillRect(10, 10 + i * 20, 10, 10);
ctx.fillStyle = '#e4e6eb';
ctx.fillText(func.name, 25, 19 + i * 20);
});
}
// Neural Network Graph
function drawNetworkGraph(ctx, canvas) {
const layers = [2, 3, 3, 1];
const width = canvas.width;
const height = canvas.height;
const layerWidth = width / (layers.length + 1);
ctx.fillStyle = 'rgba(0, 212, 255, 0.05)';
ctx.fillRect(0, 0, width, height);
// Draw neurons and connections
const neuronPositions = [];
layers.forEach((numNeurons, layerIdx) => {
const x = (layerIdx + 1) * layerWidth;
const positions = [];
for (let i = 0; i < numNeurons; i++) {
const y = height / (numNeurons + 1) * (i + 1);
positions.push({ x, y });
// Draw connections to next layer
if (layerIdx < layers.length - 1) {
const nextLayerPositions = [];
const nextX = (layerIdx + 2) * layerWidth;
for (let j = 0; j < layers[layerIdx + 1]; j++) {
const nextY = height / (layers[layerIdx + 1] + 1) * (j + 1);
nextLayerPositions.push({ x: nextX, y: nextY });
}
nextLayerPositions.forEach(next => {
ctx.strokeStyle = 'rgba(0, 212, 255, 0.2)';
ctx.lineWidth = 1;
ctx.beginPath();
ctx.moveTo(x, y);
ctx.lineTo(next.x, next.y);
ctx.stroke();
});
}
}
// Draw neurons
positions.forEach(pos => {
ctx.fillStyle = '#00d4ff';
ctx.beginPath();
ctx.arc(pos.x, pos.y, 8, 0, Math.PI * 2);
ctx.fill();
});
neuronPositions.push(positions);
});
// Labels
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 12px Arial';
ctx.textAlign = 'center';
ctx.fillText('Input', layerWidth, height - 10);
ctx.fillText('Hidden 1', layerWidth * 2, height - 10);
ctx.fillText('Hidden 2', layerWidth * 3, height - 10);
ctx.fillText('Output', layerWidth * 4, height - 10);
}
// Convolution Animation
function drawConvolutionAnimation(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
// Draw input image
ctx.fillStyle = 'rgba(0, 212, 255, 0.1)';
ctx.fillRect(20, 20, 150, 150);
ctx.strokeStyle = '#00d4ff';
ctx.lineWidth = 2;
ctx.strokeRect(20, 20, 150, 150);
// Draw filter
ctx.fillStyle = 'rgba(255, 107, 53, 0.1)';
const filterPos = 60 + Math.sin(Date.now() / 1000) * 40;
ctx.fillRect(filterPos, 60, 60, 60);
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 3;
ctx.strokeRect(filterPos, 60, 60, 60);
// Draw output
ctx.fillStyle = 'rgba(0, 255, 136, 0.1)';
ctx.fillRect(width - 170, 20, 150, 150);
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 2;
ctx.strokeRect(width - 170, 20, 150, 150);
// Draw feature map
for (let i = 0; i < 5; i++) {
for (let j = 0; j < 5; j++) {
const intensity = Math.random() * 100;
ctx.fillStyle = `rgba(0, 212, 255, ${intensity / 100})`;
ctx.fillRect(width - 160 + i * 25, 30 + j * 25, 20, 20);
}
}
// Labels
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 12px Arial';
ctx.textAlign = 'left';
ctx.fillText('Input Image', 20, 190);
ctx.fillText('Filter', filterPos, 140);
ctx.fillText('Feature Map', width - 170, 190);
}
// Loss Landscape
function drawLossLandscape(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
for (let x = 0; x < width; x += 20) {
for (let y = 0; y < height; y += 20) {
const nx = (x - width / 2) / (width / 4);
const ny = (y - height / 2) / (height / 4);
const loss = nx * nx + ny * ny;
const intensity = Math.min(255, loss * 50);
ctx.fillStyle = `rgb(${intensity}, ${100}, ${255 - intensity})`;
ctx.fillRect(x, y, 20, 20);
}
}
// Draw descent path
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 2;
ctx.beginPath();
const startX = width / 2 + 80;
const startY = height / 2 + 80;
ctx.moveTo(startX, startY);
for (let i = 0; i < 20; i++) {
const angle = Math.atan2(startY - height / 2, startX - width / 2);
const newX = startX - Math.cos(angle) * 15;
const newY = startY - Math.sin(angle) * 15;
ctx.lineTo(newX, newY);
}
ctx.stroke();
// Minimum point
ctx.fillStyle = '#00ff88';
ctx.beginPath();
ctx.arc(width / 2, height / 2, 8, 0, Math.PI * 2);
ctx.fill();
}
// YOLO Grid
function drawYOLOGrid(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
const gridSize = 7;
const cellWidth = width / gridSize;
const cellHeight = height / gridSize;
// Draw grid
ctx.strokeStyle = 'rgba(0, 212, 255, 0.3)';
ctx.lineWidth = 1;
for (let i = 0; i <= gridSize; i++) {
ctx.beginPath();
ctx.moveTo(i * cellWidth, 0);
ctx.lineTo(i * cellWidth, height);
ctx.stroke();
ctx.beginPath();
ctx.moveTo(0, i * cellHeight);
ctx.lineTo(width, i * cellHeight);
ctx.stroke();
}
// Draw detected objects
const detections = [
{ x: 2, y: 2, w: 2, h: 2, conf: 0.95 },
{ x: 4, y: 5, w: 1.5, h: 1.5, conf: 0.87 }
];
detections.forEach(det => {
ctx.fillStyle = `rgba(255, 107, 53, ${det.conf * 0.5})`;
ctx.fillRect(det.x * cellWidth, det.y * cellHeight, det.w * cellWidth, det.h * cellHeight);
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 2;
ctx.strokeRect(det.x * cellWidth, det.y * cellHeight, det.w * cellWidth, det.h * cellHeight);
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 12px Arial';
ctx.fillText((det.conf * 100).toFixed(0) + '%', det.x * cellWidth + 5, det.y * cellHeight + 15);
});
}
// Attention Matrix
function drawAttentionMatrix(ctx, canvas) {
const size = 8;
const cellSize = Math.min(canvas.width, canvas.height) / size;
for (let i = 0; i < size; i++) {
for (let j = 0; j < size; j++) {
const distance = Math.abs(i - j);
const attention = Math.exp(-distance / 2);
const intensity = Math.floor(attention * 255);
ctx.fillStyle = `rgb(${intensity}, 100, ${200 - intensity})`;
ctx.fillRect(i * cellSize, j * cellSize, cellSize, cellSize);
}
}
// Add labels
ctx.fillStyle = '#e4e6eb';
ctx.font = '10px Arial';
ctx.textAlign = 'center';
for (let i = 0; i < size; i++) {
ctx.fillText('w' + i, i * cellSize + cellSize / 2, canvas.height - 5);
}
}
// Math Visualization
function drawMathVisualization(moduleId) {
const canvas = document.getElementById(moduleId + '-math-canvas');
if (!canvas) return;
const ctx = canvas.getContext('2d');
ctx.clearRect(0, 0, canvas.width, canvas.height);
ctx.fillStyle = '#0f1419';
ctx.fillRect(0, 0, canvas.width, canvas.height);
const mathVizMap = {
'nn-basics': () => drawNNMath(ctx, canvas),
'activation': () => drawActivationDerivatives(ctx, canvas),
'loss': () => drawLossComparison(ctx, canvas),
'optimizers': () => drawOptimizerSteps(ctx, canvas),
'backprop': () => drawChainRule(ctx, canvas),
'conv-layer': () => drawConvolutionMath(ctx, canvas),
'pooling': () => drawPoolingMath(ctx, canvas),
'regularization': () => drawRegularizationMath(ctx, canvas),
'transformers': () => drawAttentionMath(ctx, canvas),
'rnn': () => drawRNNMath(ctx, canvas),
'gnn': () => drawGNNMath(ctx, canvas)
};
if (mathVizMap[moduleId]) {
mathVizMap[moduleId]();
} else {
drawDefaultMathVisualization(ctx, canvas);
}
}
// Application Visualization
function drawApplicationVisualization(moduleId) {
const canvas = document.getElementById(moduleId + '-app-canvas');
if (!canvas) return;
const ctx = canvas.getContext('2d');
ctx.clearRect(0, 0, canvas.width, canvas.height);
ctx.fillStyle = '#0f1419';
ctx.fillRect(0, 0, canvas.width, canvas.height);
const appVizMap = {
'nn-basics': () => drawNNApplications(ctx, canvas),
'cnn-basics': () => drawCNNApplications(ctx, canvas),
'conv-layer': () => drawConvolutionApplications(ctx, canvas),
'yolo': () => drawYOLOApplications(ctx, canvas),
'semantic-seg': () => drawSegmentationApplications(ctx, canvas),
'instance-seg': () => drawInstanceSegmentationApps(ctx, canvas),
'face-recog': () => drawFaceRecognitionApps(ctx, canvas),
'transformers': () => drawTransformerApps(ctx, canvas),
'bert': () => drawBERTApplications(ctx, canvas),
'gpt': () => drawGPTApplications(ctx, canvas),
'gans': () => drawGANApplications(ctx, canvas),
'diffusion': () => drawDiffusionApplications(ctx, canvas),
'gnn': () => drawGNNApplications(ctx, canvas)
};
if (appVizMap[moduleId]) {
appVizMap[moduleId]();
} else {
drawDefaultApplicationVisualization(ctx, canvas);
}
}
// Math visualization helper functions
function drawNNMath(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 18px Arial';
ctx.textAlign = 'center';
ctx.fillText('Forward Pass: y = ฯ(Wx + b)', canvas.width / 2, 50);
ctx.font = '14px Arial';
ctx.fillStyle = '#00ff88';
ctx.fillText('Linear combination + Non-linearity', canvas.width / 2, 100);
ctx.fillStyle = '#ffa500';
ctx.fillText('W: weights, b: bias, ฯ: activation', canvas.width / 2, 150);
}
function drawActivationDerivatives(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
const centerX = width / 2;
const centerY = height / 2;
const scale = 40;
ctx.strokeStyle = 'rgba(0, 212, 255, 0.2)';
ctx.lineWidth = 1;
for (let i = -5; i <= 5; i += 1) {
ctx.beginPath();
ctx.moveTo(centerX + i * scale, centerY - 5 * scale);
ctx.lineTo(centerX + i * scale, centerY + 5 * scale);
ctx.stroke();
}
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 3;
ctx.beginPath();
for (let x = -5; x <= 5; x += 0.1) {
const y = 1 / (1 + Math.exp(-x)) * (1 - 1 / (1 + Math.exp(-x)));
const canvasX = centerX + x * scale;
const canvasY = centerY - y * scale * 10;
if (x === -5) ctx.moveTo(canvasX, canvasY);
else ctx.lineTo(canvasX, canvasY);
}
ctx.stroke();
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText("Sigmoid Derivative: ฯ'(x) = ฯ(x)(1-ฯ(x))", canvas.width / 2, 30);
}
function drawLossComparison(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
// MSE
ctx.fillStyle = 'rgba(0, 212, 255, 0.2)';
ctx.fillRect(20, 60, width / 2 - 30, height - 100);
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.fillText('MSE Loss', width / 4, 45);
ctx.font = '12px Arial';
ctx.fillText('L = (1/n)ฮฃ(y-ลท)ยฒ', width / 4, 90);
ctx.fillText('Regression', width / 4, 115);
// Cross-Entropy
ctx.fillStyle = 'rgba(255, 107, 53, 0.2)';
ctx.fillRect(width / 2 + 10, 60, width / 2 - 30, height - 100);
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 14px Arial';
ctx.fillText('Cross-Entropy Loss', width * 3 / 4, 45);
ctx.font = '12px Arial';
ctx.fillText('L = -ฮฃ(yยทlog(ลท))', width * 3 / 4, 90);
ctx.fillText('Classification', width * 3 / 4, 115);
}
function drawOptimizerSteps(ctx, canvas) {
const width = canvas.width;
const height = canvas.height;
const centerY = height / 2;
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('SGD', width / 4, 50);
ctx.font = '12px Arial';
ctx.fillText('w = w - ฮฑยทโL', width / 4, 100);
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 16px Arial';
ctx.fillText('Momentum', width / 2, 50);
ctx.font = '12px Arial';
ctx.fillText('v = ฮฒยทv + (1-ฮฒ)ยทโL', width / 2, 100);
ctx.fillStyle = '#ffa500';
ctx.font = 'bold 16px Arial';
ctx.fillText('Adam', width * 3 / 4, 50);
ctx.font = '12px Arial';
ctx.fillText('Adaptive learning rate', width * 3 / 4, 100);
}
function drawChainRule(ctx, canvas) {
const width = canvas.width;
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Backpropagation Chain Rule', width / 2, 50);
ctx.font = '12px Arial';
ctx.fillStyle = '#00d4ff';
ctx.fillText('dL/dW = dL/dลท ยท dลท/da ยท da/dz ยท dz/dW', width / 2, 100);
ctx.fillStyle = '#ffa500';
ctx.fillText('Compute gradient by multiplying partial derivatives', width / 2, 150);
}
function drawConvolutionMath(ctx, canvas) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Convolution Operation', canvas.width / 2, 50);
ctx.font = '12px Arial';
ctx.fillStyle = '#00d4ff';
ctx.fillText('y[i,j] = ฮฃ ฮฃ w[m,n] * x[i+m,j+n] + b', canvas.width / 2, 100);
ctx.fillStyle = '#00ff88';
ctx.fillText('Sliding window element-wise multiplication and summation', canvas.width / 2, 150);
}
function drawPoolingMath(ctx, canvas) {
const width = canvas.width;
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Max Pooling', width / 3, 50);
ctx.font = '12px Arial';
ctx.fillText('y = max(neighborhood)', width / 3, 100);
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.fillText('Average Pooling', width * 2 / 3, 50);
ctx.font = '12px Arial';
ctx.fillText('y = avg(neighborhood)', width * 2 / 3, 100);
ctx.fillStyle = '#ffa500';
ctx.font = '11px Arial';
ctx.textAlign = 'center';
ctx.fillText('Reduces spatial dimensions', width / 2, 150);
}
function drawRegularizationMath(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('L1 Regularization: L = Loss + ฮปฮฃ|w|', canvas.width / 2, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('L2 Regularization: L = Loss + ฮปฮฃwยฒ', canvas.width / 2, 110);
ctx.fillStyle = '#ffa500';
ctx.fillText('Prevents overfitting by penalizing large weights', canvas.width / 2, 160);
}
function drawAttentionMath(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Attention Mechanism', canvas.width / 2, 50);
ctx.font = '12px Arial';
ctx.fillStyle = '#00ff88';
ctx.fillText('Attention(Q,K,V) = softmax(QK^T/โd_k) ยท V', canvas.width / 2, 100);
ctx.fillStyle = '#ffa500';
ctx.fillText('Query-Key matching determines how much to focus on each value', canvas.width / 2, 150);
}
function drawRNNMath(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('RNN Hidden State Update', canvas.width / 2, 50);
ctx.font = '12px Arial';
ctx.fillStyle = '#00ff88';
ctx.fillText('h_t = ฯ(W_hยทh_(t-1) + W_xยทx_t + b)', canvas.width / 2, 100);
ctx.fillStyle = '#ffa500';
ctx.fillText('Processes sequences step-by-step with recurrent connections', canvas.width / 2, 150);
}
// Application visualization helper functions
function drawNNApplications(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ฑ Stock Price Prediction', canvas.width / 4, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('๐ฅ Medical Diagnosis', canvas.width / 2, 60);
ctx.fillStyle = '#ffa500';
ctx.fillText('๐ฎ Game AI', canvas.width * 3 / 4, 60);
ctx.fillStyle = '#ff6b35';
ctx.font = '12px Arial';
ctx.fillText('Fraud Detection', canvas.width / 4, 120);
ctx.fillStyle = '#00d4ff';
ctx.fillText('Recommendation Systems', canvas.width / 2, 120);
ctx.fillStyle = '#00ff88';
ctx.fillText('Credit Scoring', canvas.width * 3 / 4, 120);
}
function drawCNNApplications(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Image Classification', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('Object Detection', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Deep Learning Backbone', canvas.width / 2, 150);
}
function drawConvolutionApplications(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ท Image Feature Extraction', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('๐ Edge Detection', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Foundation of Computer Vision', canvas.width / 2, 150);
}
function drawYOLOApplications(ctx, canvas) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Autonomous Driving', canvas.width / 3, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐น Real-time Video Detection', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.font = '12px Arial';
ctx.fillText('Ultra-fast inference for live applications', canvas.width / 2, 150);
}
function drawSegmentationApplications(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ฅ Medical Imaging', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('๐ Autonomous Vehicles', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Pixel-level understanding of scenes', canvas.width / 2, 150);
}
function drawInstanceSegmentationApps(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ฅ Person Detection & Tracking', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('๐ Object Instance Counting', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Separates overlapping objects', canvas.width / 2, 150);
}
function drawFaceRecognitionApps(ctx, canvas) {
ctx.fillStyle = '#ffa500';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ฑ Phone Unlock', canvas.width / 3, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐ Security Systems', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.font = '12px Arial';
ctx.fillText('Identity verification and access control', canvas.width / 2, 150);
}
function drawTransformerApps(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ฌ ChatGPT / LLMs', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('๐ Machine Translation', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Foundation of modern NLP and beyond', canvas.width / 2, 150);
}
function drawBERTApplications(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Semantic Search', canvas.width / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.fillText('โ Question Answering', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#ffa500';
ctx.font = '12px Arial';
ctx.fillText('Deep language understanding', canvas.width / 2, 150);
}
function drawGPTApplications(ctx, canvas) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('โ๏ธ Text Generation', canvas.width / 3, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐ก Idea Assistance', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.font = '12px Arial';
ctx.fillText('Powerful autoregressive language models', canvas.width / 2, 150);
}
function drawGANApplications(ctx, canvas) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐จ Image Generation', canvas.width / 3, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐ญ Style Transfer', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.font = '12px Arial';
ctx.fillText('Creative content generation and enhancement', canvas.width / 2, 150);
}
function drawDiffusionApplications(ctx, canvas) {
ctx.fillStyle = '#ffa500';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ผ๏ธ Image Synthesis', canvas.width / 3, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐ฌ Stable Diffusion', canvas.width * 2 / 3, 60);
ctx.fillStyle = '#00ff88';
ctx.font = '12px Arial';
ctx.fillText('State-of-the-art generative AI', canvas.width / 2, 150);
}
// Missing visualization stub functions
function drawNeuronAnimation(ctx, canvas) {
drawNetworkGraph(ctx, canvas);
}
function drawDecisionBoundary(ctx, canvas) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
// Draw decision boundary line
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 3;
ctx.beginPath();
ctx.moveTo(0, centerY);
ctx.lineTo(canvas.width, centerY);
ctx.stroke();
// Draw sample points
for (let i = 0; i < 20; i++) {
const x = Math.random() * canvas.width;
const y = Math.random() * canvas.height;
ctx.fillStyle = y < centerY ? '#00d4ff' : '#00ff88';
ctx.beginPath();
ctx.arc(x, y, 5, 0, Math.PI * 2);
ctx.fill();
}
}
function drawWeightDistribution(ctx, canvas) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
// Draw Gaussian distribution
ctx.strokeStyle = '#00d4ff';
ctx.lineWidth = 2;
ctx.beginPath();
for (let x = -100; x <= 100; x += 2) {
const y = Math.exp(-(x * x) / 500) * 80;
const canvasX = centerX + x;
const canvasY = centerY - y;
if (x === -100) ctx.moveTo(canvasX, canvasY);
else ctx.lineTo(canvasX, canvasY);
}
ctx.stroke();
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Weight Distribution (Xavier/He Init)', centerX, 50);
}
function drawConvergencePaths(ctx, canvas) {
drawLossLandscape(ctx, canvas);
}
function drawGradientFlow(ctx, canvas) {
drawChainRule(ctx, canvas);
}
function drawOverfitComparison(ctx, canvas) {
const width = canvas.width;
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Without Regularization', width / 4, 40);
ctx.fillStyle = '#ff6b35';
ctx.fillText('With Regularization', width * 3 / 4, 40);
// Draw wavy overfit line
ctx.strokeStyle = '#00d4ff';
ctx.lineWidth = 2;
ctx.beginPath();
for (let x = 0; x < width / 2 - 20; x += 5) {
const y = 100 + Math.sin(x / 10) * 30 + Math.random() * 20;
if (x === 0) ctx.moveTo(x + 20, y);
else ctx.lineTo(x + 20, y);
}
ctx.stroke();
// Draw smooth regularized line
ctx.strokeStyle = '#ff6b35';
ctx.beginPath();
for (let x = 0; x < width / 2 - 20; x += 5) {
const y = 100 + Math.sin(x / 20) * 15;
if (x === 0) ctx.moveTo(x + width / 2 + 20, y);
else ctx.lineTo(x + width / 2 + 20, y);
}
ctx.stroke();
}
function drawBatchNormalization(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Batch Normalization: ฮผ=0, ฯยฒ=1', canvas.width / 2, 50);
// Draw before/after distributions
ctx.fillStyle = '#ffa500';
ctx.fillText('Input Distribution', canvas.width / 4, 100);
ctx.fillStyle = '#00ff88';
ctx.fillText('Normalized Distribution', canvas.width * 3 / 4, 100);
}
function drawImageMatrix(ctx, canvas) {
const cellSize = 20;
for (let i = 0; i < 10; i++) {
for (let j = 0; j < 10; j++) {
const intensity = Math.random();
ctx.fillStyle = `rgba(0, 212, 255, ${intensity})`;
ctx.fillRect(i * cellSize + 100, j * cellSize + 100, cellSize, cellSize);
}
}
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Image as Matrix (Pixel Values)', canvas.width / 2, 50);
}
function drawPoolingDemo(ctx, canvas) {
const cellSize = 30;
const input = [[20, 30, 0, 5], [8, 12, 2, 0], [34, 70, 37, 4], [112, 100, 25, 12]];
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Max Pooling Demo (2x2)', canvas.width / 2, 30);
// Draw input matrix
for (let i = 0; i < 4; i++) {
for (let j = 0; j < 4; j++) {
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(50 + j * cellSize, 50 + i * cellSize, cellSize, cellSize);
ctx.fillStyle = '#e4e6eb';
ctx.font = '10px Arial';
ctx.fillText(input[i][j], 50 + j * cellSize + cellSize / 2, 50 + i * cellSize + cellSize / 2 + 4);
}
}
// Draw output (max pooled)
const pooled = [[20, 30], [112, 37]];
for (let i = 0; i < 2; i++) {
for (let j = 0; j < 2; j++) {
ctx.strokeStyle = '#00ff88';
ctx.strokeRect(250 + j * cellSize * 1.5, 70 + i * cellSize * 1.5, cellSize * 1.5, cellSize * 1.5);
ctx.fillStyle = '#00ff88';
ctx.font = 'bold 12px Arial';
ctx.fillText(pooled[i][j], 250 + j * cellSize * 1.5 + cellSize * 0.75, 70 + i * cellSize * 1.5 + cellSize * 0.75 + 5);
}
}
}
function drawCNNArchitecture(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 12px Arial';
ctx.textAlign = 'center';
ctx.fillText('Input', 60, 200);
ctx.fillText('Conv', 160, 200);
ctx.fillText('Pool', 260, 200);
ctx.fillText('Conv', 360, 200);
ctx.fillText('Pool', 460, 200);
ctx.fillText('FC', 560, 200);
ctx.fillText('Output', 660, 200);
// Draw blocks
const blocks = [60, 160, 260, 360, 460, 560, 660];
blocks.forEach((x, i) => {
const height = i === 0 ? 100 : (i < blocks.length - 2 ? 80 - i * 10 : 60);
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(x - 30, 100, 60, height);
});
}
function drawLearnedFilters(ctx, canvas) {
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('CNN Learned Filters', canvas.width / 2, 30);
const labels = ['Edges', 'Textures', 'Patterns', 'Objects'];
labels.forEach((label, i) => {
const x = (i + 1) * canvas.width / 5;
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 12px Arial';
ctx.fillText(label, x, 80);
// Draw filter representation
for (let j = 0; j < 3; j++) {
for (let k = 0; k < 3; k++) {
const intensity = Math.random();
ctx.fillStyle = `rgba(0, 212, 255, ${intensity})`;
ctx.fillRect(x - 20 + k * 12, 100 + j * 12, 10, 10);
}
}
});
}
function drawLeNetArchitecture(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawAlexNetArchitecture(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawVGGArchitecture(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawResNetArchitecture(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawInceptionModule(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawMobileNetArchitecture(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawTransferLearning(ctx, canvas) { drawCNNArchitecture(ctx, canvas); }
function drawBoundingBoxes(ctx, canvas) {
// Draw sample image
ctx.fillStyle = 'rgba(0, 212, 255, 0.1)';
ctx.fillRect(50, 50, 300, 300);
// Draw bounding boxes
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 3;
ctx.strokeRect(100, 100, 150, 150);
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 12px Arial';
ctx.fillText('Dog 95%', 105, 95);
ctx.strokeStyle = '#00ff88';
ctx.strokeRect(180, 200, 100, 80);
ctx.fillStyle = '#00ff88';
ctx.fillText('Cat 87%', 185, 195);
}
function drawRCNNPipeline(ctx, canvas) { drawBoundingBoxes(ctx, canvas); }
function drawSSDDetector(ctx, canvas) { drawBoundingBoxes(ctx, canvas); }
function drawSemanticSegmentation(ctx, canvas) {
const cellSize = 15;
const colors = ['rgba(0, 212, 255, 0.5)', 'rgba(255, 107, 53, 0.5)', 'rgba(0, 255, 136, 0.5)'];
for (let i = 0; i < 20; i++) {
for (let j = 0; j < 20; j++) {
ctx.fillStyle = colors[Math.floor(Math.random() * colors.length)];
ctx.fillRect(i * cellSize + 100, j * cellSize + 50, cellSize, cellSize);
}
}
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Pixel-wise Classification', canvas.width / 2, 30);
}
function drawInstanceSegmentation(ctx, canvas) { drawSemanticSegmentation(ctx, canvas); }
function drawFaceEmbeddings(ctx, canvas) {
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Face Embedding Space', canvas.width / 2, 30);
// Draw embedding vectors
const faces = 5;
for (let i = 0; i < faces; i++) {
const x = 100 + Math.random() * (canvas.width - 200);
const y = 100 + Math.random() * 200;
ctx.fillStyle = '#00d4ff';
ctx.beginPath();
ctx.arc(x, y, 10, 0, Math.PI * 2);
ctx.fill();
}
}
function drawAutoencoderArchitecture(ctx, canvas) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 12px Arial';
ctx.textAlign = 'center';
const stages = ['Input', 'Encoder', 'Latent', 'Decoder', 'Output'];
stages.forEach((label, i) => {
const x = (i + 1) * canvas.width / 6;
ctx.fillText(label, x, 50);
const height = i === 2 ? 40 : (i === 0 || i === 4 ? 100 : 70);
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(x - 30, 100, 60, height);
});
}
function drawGANsGame(ctx, canvas) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Generator', canvas.width / 3, 50);
ctx.fillStyle = '#00d4ff';
ctx.fillText('Discriminator', canvas.width * 2 / 3, 50);
// DrawGenerator
ctx.strokeStyle = '#ff6b35';
ctx.strokeRect(canvas.width / 3 - 50, 100, 100, 100);
// Draw Discriminator
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(canvas.width * 2 / 3 - 50, 100, 100, 100);
// Draw arrow
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(canvas.width / 3 + 50, 150);
ctx.lineTo(canvas.width * 2 / 3 - 50, 150);
ctx.stroke();
}
function drawDiffusionProcess(ctx, canvas) {
const steps = 5;
const stepWidth = canvas.width / (steps + 1);
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Diffusion Process: From Noise to Image', canvas.width / 2, 30);
for (let i = 0; i < steps; i++) {
const x = (i + 1) * stepWidth;
const noise = 1 - (i / steps);
ctx.fillStyle = `rgba(0, 212, 255, ${1 - noise})`;
ctx.fillRect(x - 40, 100, 80, 80);
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(x - 40, 100, 80, 80);
}
}
function drawRNNUnrolled(ctx, canvas) {
const cells = 5;
const cellWidth = canvas.width / (cells + 1);
ctx.fillStyle = '#e4e6eb';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('Unrolled RNN', canvas.width / 2, 30);
for (let i = 0; i < cells; i++) {
const x = (i + 1) * cellWidth;
ctx.strokeStyle = '#00d4ff';
ctx.strokeRect(x - 30, 100, 60, 60);
if (i < cells - 1) {
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(x + 30, 130);
ctx.lineTo(x + cellWidth - 30, 130);
ctx.stroke();
}
}
}
function drawBERTProcess(ctx, canvas) { drawAttentionMatrix(ctx, canvas); }
function drawGPTGeneration(ctx, canvas) { drawAttentionMatrix(ctx, canvas); }
function drawVisionTransformer(ctx, canvas) { drawAttentionMatrix(ctx, canvas); }
function drawVisualization(moduleId) {
drawConceptsVisualization(moduleId);
}
// Animation and download utilities
let animationFrameId = null;
function toggleVizAnimation(moduleId) {
const btn = event.target;
window.vizAnimating = !window.vizAnimating;
if (window.vizAnimating) {
btn.textContent = 'โน๏ธ Stop';
btn.style.background = 'linear-gradient(135deg, #ff4444, #cc0000)';
animateVisualization(moduleId);
} else {
btn.textContent = 'โถ๏ธ Animate';
btn.style.background = '';
if (animationFrameId) {
cancelAnimationFrame(animationFrameId);
animationFrameId = null;
}
}
}
function animateVisualization(moduleId) {
if (!window.vizAnimating) return;
const canvas = document.getElementById(moduleId + '-canvas');
if (!canvas) return;
const ctx = canvas.getContext('2d');
ctx.clearRect(0, 0, canvas.width, canvas.height);
ctx.fillStyle = '#0f1419';
ctx.fillRect(0, 0, canvas.width, canvas.height);
// Call the appropriate animated drawing function
const animatedVizMap = {
'nn-basics': drawAnimatedNetwork,
'perceptron': drawAnimatedDecisionBoundary,
'mlp': drawAnimatedMLP,
'activation': drawAnimatedActivations,
'conv-layer': drawAnimatedConvolution,
'gnn': drawAnimatedGNN,
'transformers': drawAnimatedAttention,
'backprop': drawAnimatedGradientFlow,
'gans': drawAnimatedGAN,
'diffusion': drawAnimatedDiffusion,
'rnn': drawAnimatedRNN
};
if (animatedVizMap[moduleId]) {
animatedVizMap[moduleId](ctx, canvas, Date.now());
} else {
// Default animation - pulsing visualization
drawDefaultAnimation(ctx, canvas, Date.now());
}
animationFrameId = requestAnimationFrame(() => animateVisualization(moduleId));
}
// Default animation for modules without specific animations
function drawDefaultAnimation(ctx, canvas, time) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
const pulse = Math.sin(time / 300) * 0.3 + 0.7;
// Animated neural network
const layers = [3, 4, 4, 2];
const layerWidth = canvas.width / (layers.length + 1);
layers.forEach((neurons, layerIdx) => {
const x = (layerIdx + 1) * layerWidth;
const layerHeight = canvas.height / (neurons + 1);
for (let i = 0; i < neurons; i++) {
const y = (i + 1) * layerHeight;
const radius = 12 + Math.sin(time / 200 + layerIdx + i) * 3;
// Draw neuron
ctx.fillStyle = `rgba(0, 212, 255, ${pulse})`;
ctx.beginPath();
ctx.arc(x, y, radius, 0, Math.PI * 2);
ctx.fill();
// Draw connections to next layer
if (layerIdx < layers.length - 1) {
const nextLayerHeight = canvas.height / (layers[layerIdx + 1] + 1);
const nextX = (layerIdx + 2) * layerWidth;
for (let j = 0; j < layers[layerIdx + 1]; j++) {
const nextY = (j + 1) * nextLayerHeight;
const signalProgress = ((time / 500) + layerIdx * 0.5) % 1;
ctx.strokeStyle = `rgba(0, 212, 255, ${0.3 + signalProgress * 0.3})`;
ctx.lineWidth = 1;
ctx.beginPath();
ctx.moveTo(x + radius, y);
ctx.lineTo(nextX - 12, nextY);
ctx.stroke();
// Animated signal dot
const dotX = x + radius + (nextX - 12 - x - radius) * signalProgress;
const dotY = y + (nextY - y) * signalProgress;
ctx.fillStyle = '#00ff88';
ctx.beginPath();
ctx.arc(dotX, dotY, 3, 0, Math.PI * 2);
ctx.fill();
}
}
}
});
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Neural Network Animation', centerX, 25);
}
// Animated GNN with message passing
function drawAnimatedGNN(ctx, canvas, time) {
ctx.fillStyle = '#9900ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Graph Neural Network - Message Passing', canvas.width / 2, 30);
const nodes = [
{ x: 100, y: 100 }, { x: 200, y: 60 }, { x: 320, y: 120 },
{ x: 150, y: 200 }, { x: 400, y: 80 }, { x: 450, y: 180 }
];
const edges = [[0, 1], [0, 3], [1, 2], [1, 4], [2, 3], [2, 4], [4, 5]];
// Draw edges
ctx.strokeStyle = 'rgba(153, 0, 255, 0.4)';
ctx.lineWidth = 2;
edges.forEach(e => {
ctx.beginPath();
ctx.moveTo(nodes[e[0]].x, nodes[e[0]].y);
ctx.lineTo(nodes[e[1]].x, nodes[e[1]].y);
ctx.stroke();
});
// Draw animated message passing
const messageProgress = (time / 1000) % 1;
ctx.fillStyle = '#00ff88';
edges.forEach((e, idx) => {
const progress = (messageProgress + idx * 0.15) % 1;
const x = nodes[e[0]].x + (nodes[e[1]].x - nodes[e[0]].x) * progress;
const y = nodes[e[0]].y + (nodes[e[1]].y - nodes[e[0]].y) * progress;
ctx.beginPath();
ctx.arc(x, y, 5, 0, Math.PI * 2);
ctx.fill();
});
// Draw nodes with pulse
const pulse = Math.sin(time / 300) * 5 + 15;
nodes.forEach((n, i) => {
ctx.fillStyle = '#9900ff';
ctx.beginPath();
ctx.arc(n.x, n.y, pulse, 0, Math.PI * 2);
ctx.fill();
ctx.fillStyle = 'white';
ctx.font = '12px Arial';
ctx.textAlign = 'center';
ctx.fillText(i, n.x, n.y + 4);
});
}
// Animated attention matrix
function drawAnimatedAttention(ctx, canvas, time) {
const words = ['The', 'cat', 'sat', 'on', 'mat'];
const cellSize = 50;
const startX = (canvas.width - words.length * cellSize) / 2;
const startY = 80;
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Self-Attention Animation', canvas.width / 2, 30);
// Draw words
ctx.font = '12px Arial';
words.forEach((word, i) => {
ctx.fillStyle = '#e4e6eb';
ctx.fillText(word, startX + i * cellSize + cellSize / 2, startY - 10);
ctx.save();
ctx.translate(startX - 20, startY + i * cellSize + cellSize / 2);
ctx.fillText(word, 0, 0);
ctx.restore();
});
// Animated attention weights
for (let i = 0; i < words.length; i++) {
for (let j = 0; j < words.length; j++) {
const baseWeight = i === j ? 0.8 : 0.2 + Math.abs(i - j) * 0.1;
const animatedWeight = baseWeight + Math.sin(time / 500 + i + j) * 0.2;
const alpha = Math.max(0.1, Math.min(1, animatedWeight));
ctx.fillStyle = `rgba(0, 212, 255, ${alpha})`;
ctx.fillRect(startX + j * cellSize + 2, startY + i * cellSize + 2, cellSize - 4, cellSize - 4);
ctx.fillStyle = '#e4e6eb';
ctx.font = '10px Arial';
ctx.fillText(animatedWeight.toFixed(2), startX + j * cellSize + cellSize / 2, startY + i * cellSize + cellSize / 2 + 4);
}
}
}
// Animated gradient flow for backprop
function drawAnimatedGradientFlow(ctx, canvas, time) {
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Backpropagation - Gradient Flow', canvas.width / 2, 30);
const layers = [2, 4, 4, 1];
const layerWidth = canvas.width / (layers.length + 1);
// Forward pass (left to right) - blue
const forwardProgress = (time / 2000) % 1;
layers.forEach((neurons, layerIdx) => {
const x = (layerIdx + 1) * layerWidth;
const layerHeight = canvas.height / (neurons + 1);
for (let i = 0; i < neurons; i++) {
const y = (i + 1) * layerHeight;
// Pulse effect based on forward pass
const isActive = forwardProgress > layerIdx / layers.length;
const radius = isActive ? 15 + Math.sin(time / 200) * 3 : 12;
ctx.fillStyle = isActive ? '#00d4ff' : 'rgba(0, 212, 255, 0.3)';
ctx.beginPath();
ctx.arc(x, y, radius, 0, Math.PI * 2);
ctx.fill();
}
});
// Backward pass (right to left) - orange/red gradients
const backwardProgress = ((time / 2000) + 0.5) % 1;
for (let layerIdx = layers.length - 2; layerIdx >= 0; layerIdx--) {
const x1 = (layerIdx + 1) * layerWidth;
const x2 = (layerIdx + 2) * layerWidth;
const gradientActive = backwardProgress > (layers.length - 2 - layerIdx) / (layers.length - 1);
if (gradientActive) {
const gradX = x2 - (x2 - x1) * ((backwardProgress * (layers.length - 1)) % 1);
ctx.fillStyle = '#ff6b35';
ctx.beginPath();
ctx.arc(gradX, canvas.height / 2, 8, 0, Math.PI * 2);
ctx.fill();
}
}
ctx.fillStyle = '#e4e6eb';
ctx.font = '12px Arial';
ctx.fillText('Forward: Blue โ | Backward: Orange โ', canvas.width / 2, canvas.height - 20);
}
// Animated network for nn-basics
function drawAnimatedNetwork(ctx, canvas, time) {
drawDefaultAnimation(ctx, canvas, time);
}
// Animated decision boundary for perceptron
function drawAnimatedDecisionBoundary(ctx, canvas, time) {
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
ctx.fillStyle = '#ff6b35';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Perceptron Decision Boundary', canvas.width / 2, 30);
// Animated rotating decision boundary
const angle = time / 2000;
const length = 200;
ctx.strokeStyle = '#ff6b35';
ctx.lineWidth = 3;
ctx.beginPath();
ctx.moveTo(centerX - Math.cos(angle) * length, centerY - Math.sin(angle) * length);
ctx.lineTo(centerX + Math.cos(angle) * length, centerY + Math.sin(angle) * length);
ctx.stroke();
// Fixed sample points
const points = [
{ x: 100, y: 80, c: 1 }, { x: 150, y: 100, c: 1 }, { x: 120, y: 150, c: 1 },
{ x: 400, y: 200, c: 0 }, { x: 450, y: 180, c: 0 }, { x: 380, y: 250, c: 0 }
];
points.forEach(p => {
ctx.fillStyle = p.c === 1 ? '#00d4ff' : '#00ff88';
ctx.beginPath();
ctx.arc(p.x, p.y, 8, 0, Math.PI * 2);
ctx.fill();
});
}
function drawAnimatedMLP(ctx, canvas, time) {
drawDefaultAnimation(ctx, canvas, time);
}
function drawAnimatedActivations(ctx, canvas, time) {
drawActivationFunctions(ctx, canvas);
// Add animated input marker
const x = Math.sin(time / 500) * 4;
const centerX = canvas.width / 2;
const centerY = canvas.height / 2;
const scale = 40;
ctx.fillStyle = '#ffffff';
ctx.beginPath();
ctx.arc(centerX + x * scale, centerY, 6, 0, Math.PI * 2);
ctx.fill();
ctx.strokeStyle = '#ffffff';
ctx.setLineDash([5, 5]);
ctx.beginPath();
ctx.moveTo(centerX + x * scale, 0);
ctx.lineTo(centerX + x * scale, canvas.height);
ctx.stroke();
ctx.setLineDash([]);
}
function drawAnimatedConvolution(ctx, canvas, time) {
drawConvolutionAnimation(ctx, canvas);
}
function drawAnimatedGAN(ctx, canvas, time) {
ctx.fillStyle = '#ffaa00';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('GAN Training Animation', canvas.width / 2, 30);
const phase = Math.floor(time / 1000) % 4;
// Generator
ctx.fillStyle = phase <= 1 ? '#00ff88' : 'rgba(0, 255, 136, 0.3)';
ctx.fillRect(50, 100, 100, 80);
ctx.fillStyle = '#e4e6eb';
ctx.font = '12px Arial';
ctx.fillText('Generator', 100, 145);
// Fake image
const noiseToFake = Math.sin(time / 300) * 0.5 + 0.5;
ctx.fillStyle = `rgba(255, 170, 0, ${noiseToFake})`;
ctx.fillRect(200, 110, 60, 60);
ctx.fillStyle = '#e4e6eb';
ctx.fillText('Fake', 230, 200);
// Discriminator
ctx.fillStyle = phase >= 2 ? '#ff6b35' : 'rgba(255, 107, 53, 0.3)';
ctx.fillRect(320, 100, 100, 80);
ctx.fillStyle = '#e4e6eb';
ctx.fillText('Discriminator', 370, 145);
// Output
const output = phase === 3 ? 'Real?' : 'Fake?';
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 14px Arial';
ctx.fillText(output, 370, 220);
// Arrows
ctx.strokeStyle = '#e4e6eb';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(150, 140);
ctx.lineTo(200, 140);
ctx.stroke();
ctx.beginPath();
ctx.moveTo(260, 140);
ctx.lineTo(320, 140);
ctx.stroke();
}
function drawAnimatedDiffusion(ctx, canvas, time) {
ctx.fillStyle = '#9900ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Diffusion Process Animation', canvas.width / 2, 30);
const steps = 5;
const stepWidth = canvas.width / (steps + 1);
const progress = (time / 3000) % 1;
const currentStep = Math.floor(progress * steps);
for (let i = 0; i < steps; i++) {
const x = (i + 1) * stepWidth;
const y = 150;
const noiseLevel = i / (steps - 1);
const isActive = i <= currentStep;
// Draw square with noise
ctx.fillStyle = isActive ? '#9900ff' : 'rgba(153, 0, 255, 0.3)';
ctx.fillRect(x - 30, y - 30, 60, 60);
// Add noise dots
if (noiseLevel > 0) {
for (let j = 0; j < noiseLevel * 20; j++) {
const nx = x - 25 + Math.random() * 50;
const ny = y - 25 + Math.random() * 50;
ctx.fillStyle = 'rgba(255, 255, 255, 0.5)';
ctx.fillRect(nx, ny, 2, 2);
}
}
ctx.fillStyle = '#e4e6eb';
ctx.font = '10px Arial';
ctx.fillText(`t=${i}`, x, y + 50);
}
ctx.fillStyle = '#e4e6eb';
ctx.font = '12px Arial';
ctx.fillText('Clean โ Noisy (Forward) | Noisy โ Clean (Reverse)', canvas.width / 2, canvas.height - 20);
}
function drawAnimatedRNN(ctx, canvas, time) {
ctx.fillStyle = '#00d4ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('RNN Unrolled Through Time', canvas.width / 2, 30);
const steps = 5;
const stepWidth = canvas.width / (steps + 1);
const progress = (time / 500) % steps;
const activeStep = Math.floor(progress);
for (let i = 0; i < steps; i++) {
const x = (i + 1) * stepWidth;
const y = 150;
const isActive = i === activeStep;
// Hidden state
ctx.fillStyle = isActive ? '#00d4ff' : 'rgba(0, 212, 255, 0.3)';
ctx.beginPath();
ctx.arc(x, y, 25, 0, Math.PI * 2);
ctx.fill();
ctx.fillStyle = '#e4e6eb';
ctx.font = '10px Arial';
ctx.fillText(`h${i}`, x, y + 4);
// Input arrow
ctx.strokeStyle = isActive ? '#00ff88' : 'rgba(0, 255, 136, 0.3)';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(x, y + 60);
ctx.lineTo(x, y + 25);
ctx.stroke();
ctx.fillText(`x${i}`, x, y + 75);
// Recurrent connection
if (i < steps - 1) {
ctx.strokeStyle = isActive ? '#ff6b35' : 'rgba(255, 107, 53, 0.3)';
ctx.beginPath();
ctx.moveTo(x + 25, y);
ctx.lineTo(x + stepWidth - 25, y);
ctx.stroke();
// Animated signal
if (isActive) {
const signalX = x + 25 + (stepWidth - 50) * (progress % 1);
ctx.fillStyle = '#ff6b35';
ctx.beginPath();
ctx.arc(signalX, y, 5, 0, Math.PI * 2);
ctx.fill();
}
}
}
}
function downloadViz(moduleId) {
const canvas = document.getElementById(moduleId + '-canvas');
if (!canvas) return;
const link = document.createElement('a');
link.href = canvas.toDataURL('image/png');
link.download = moduleId + '-visualization.png';
link.click();
}
function drawGraphNetwork(ctx, canvas) {
ctx.fillStyle = '#9900ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Graph Structure & Message Passing', canvas.width / 2, 30);
const nodes = [
{ x: 100, y: 100 }, { x: 200, y: 50 }, { x: 300, y: 150 },
{ x: 150, y: 250 }, { x: 400, y: 100 }, { x: 500, y: 200 }
];
const edges = [
[0, 1], [0, 3], [1, 2], [1, 4], [2, 3], [2, 4], [4, 5]
];
// Draw edges
ctx.strokeStyle = 'rgba(153, 0, 255, 0.4)';
ctx.lineWidth = 2;
edges.forEach(e => {
ctx.beginPath();
ctx.moveTo(nodes[e[0]].x, nodes[e[0]].y);
ctx.lineTo(nodes[e[1]].x, nodes[e[1]].y);
ctx.stroke();
});
// Draw nodes
nodes.forEach((n, i) => {
ctx.fillStyle = '#9900ff';
ctx.beginPath();
ctx.arc(n.x, n.y, 15, 0, Math.PI * 2);
ctx.fill();
ctx.fillStyle = 'white';
ctx.font = '12px Arial';
ctx.fillText(i, n.x, n.y + 4);
});
// Draw Message Passing Animation (fake)
const t = (Date.now() / 1000) % 2;
if (t > 1) {
ctx.strokeStyle = '#00ff88';
ctx.lineWidth = 4;
edges.forEach((e, idx) => {
if (idx % 2 === 0) {
ctx.beginPath();
ctx.moveTo(nodes[e[0]].x, nodes[e[0]].y);
ctx.lineTo(nodes[e[1]].x, nodes[e[1]].y);
ctx.stroke();
}
});
}
}
function drawGNNMath(ctx, canvas) {
ctx.fillStyle = '#9900ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Graph Convolution Math', canvas.width / 2, 50);
ctx.fillStyle = '#e4e6eb';
ctx.font = '14px Courier New';
ctx.fillText('H(l+1) = ฯ(D^-ยฝ A D^-ยฝ H(l) W(l))', canvas.width / 2, 100);
ctx.fillStyle = '#00ff88';
ctx.fillText('A = Neighborhood Connections', canvas.width / 2, 150);
ctx.fillStyle = '#ff6b35';
ctx.fillText('D = Normalization Factor', canvas.width / 2, 180);
}
function drawGNNApplications(ctx, canvas) {
ctx.fillStyle = '#9900ff';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('๐ Drug Discovery (Molecular Graphs)', canvas.width / 2, 60);
ctx.fillStyle = '#00d4ff';
ctx.fillText('๐ Traffic Flow Prediction', canvas.width / 2, 120);
ctx.fillStyle = '#ff6b35';
ctx.fillText('๐ Pinterest/Amazon Recommendations', canvas.width / 2, 180);
}
// ============ GenAI Visualizations ============
function drawVectorSpace(ctx, canvas) {
const w = canvas.width, h = canvas.height;
ctx.fillStyle = '#00c9a7';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('Vector Space โ Similarity Search', w / 2, 30);
// Draw axes
ctx.strokeStyle = 'rgba(0, 201, 167, 0.3)';
ctx.lineWidth = 1;
ctx.beginPath(); ctx.moveTo(60, h - 40); ctx.lineTo(w - 20, h - 40); ctx.stroke();
ctx.beginPath(); ctx.moveTo(60, h - 40); ctx.lineTo(60, 50); ctx.stroke();
ctx.fillStyle = '#b0b7c3'; ctx.font = '11px Arial';
ctx.fillText('Dimension 1', w / 2, h - 15);
ctx.save(); ctx.translate(15, h / 2); ctx.rotate(-Math.PI / 2);
ctx.fillText('Dimension 2', 0, 0); ctx.restore();
// Cluster A (documents)
const clusterA = [
{ x: 200, y: 120, label: 'doc1' }, { x: 230, y: 100, label: 'doc2' },
{ x: 180, y: 140, label: 'doc3' }, { x: 220, y: 150, label: 'doc4' }
];
// Cluster B
const clusterB = [
{ x: 420, y: 200, label: 'doc5' }, { x: 450, y: 220, label: 'doc6' },
{ x: 400, y: 240, label: 'doc7' }
];
// Query
const query = { x: 190, y: 130, label: '๐ Query' };
// Draw cluster backgrounds
ctx.fillStyle = 'rgba(0, 136, 255, 0.08)';
ctx.beginPath(); ctx.arc(210, 130, 60, 0, Math.PI * 2); ctx.fill();
ctx.fillStyle = 'rgba(255, 107, 53, 0.08)';
ctx.beginPath(); ctx.arc(423, 220, 55, 0, Math.PI * 2); ctx.fill();
// Draw similarity lines from query to nearest
ctx.strokeStyle = 'rgba(0, 255, 136, 0.5)';
ctx.lineWidth = 2;
ctx.setLineDash([5, 3]);
clusterA.forEach(d => {
ctx.beginPath(); ctx.moveTo(query.x, query.y);
ctx.lineTo(d.x, d.y); ctx.stroke();
});
ctx.setLineDash([]);
// Draw document points
clusterA.forEach(d => {
ctx.fillStyle = '#0088ff';
ctx.beginPath(); ctx.arc(d.x, d.y, 8, 0, Math.PI * 2); ctx.fill();
ctx.fillStyle = '#e4e6eb'; ctx.font = '10px Arial';
ctx.fillText(d.label, d.x + 12, d.y + 4);
});
clusterB.forEach(d => {
ctx.fillStyle = '#ff6b35';
ctx.beginPath(); ctx.arc(d.x, d.y, 8, 0, Math.PI * 2); ctx.fill();
ctx.fillStyle = '#e4e6eb'; ctx.font = '10px Arial';
ctx.fillText(d.label, d.x + 12, d.y + 4);
});
// Draw query
ctx.fillStyle = '#00ff88';
ctx.beginPath(); ctx.arc(query.x, query.y, 10, 0, Math.PI * 2); ctx.fill();
ctx.fillStyle = '#00ff88'; ctx.font = 'bold 12px Arial';
ctx.fillText(query.label, query.x - 15, query.y - 18);
// Legend
ctx.font = '11px Arial'; let ly = h - 80;
[['#0088ff', 'Cluster A (nearest)'], ['#ff6b35', 'Cluster B'], ['#00ff88', 'Query Vector']].forEach(([c, t]) => {
ctx.fillStyle = c; ctx.fillRect(w - 160, ly, 10, 10);
ctx.fillStyle = '#e4e6eb'; ctx.fillText(t, w - 145, ly + 9); ly += 18;
});
// Cosine score
ctx.fillStyle = '#00c9a7'; ctx.font = 'bold 13px Arial';
ctx.fillText('cos(query, doc1) = 0.97', w / 2, h - 55);
}
function drawRAGPipeline(ctx, canvas) {
const w = canvas.width, h = canvas.height;
ctx.fillStyle = '#00c9a7';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('RAG Pipeline โ Retrieval-Augmented Generation', w / 2, 30);
const boxes = [
{ x: 30, y: 70, w: 90, h: 50, label: '๐ Docs', color: '#0088ff' },
{ x: 145, y: 70, w: 90, h: 50, label: 'โ๏ธ Chunk', color: '#0088ff' },
{ x: 260, y: 70, w: 90, h: 50, label: '๐งฎ Embed', color: '#0088ff' },
{ x: 375, y: 70, w: 100, h: 50, label: '๐งฒ Vector DB', color: '#0088ff' },
{ x: 30, y: 180, w: 90, h: 50, label: 'โ Query', color: '#00ff88' },
{ x: 145, y: 180, w: 90, h: 50, label: '๐งฎ Embed', color: '#00ff88' },
{ x: 260, y: 180, w: 90, h: 50, label: '๐ Retrieve', color: '#ff6b35' },
{ x: 375, y: 180, w: 100, h: 50, label: '๐ค LLM', color: '#ffaa00' },
{ x: 500, y: 180, w: 90, h: 50, label: 'โ
Answer', color: '#00c9a7' }
];
// Draw arrows (indexing pipeline)
ctx.strokeStyle = 'rgba(0, 136, 255, 0.6)'; ctx.lineWidth = 2;
for (let i = 0; i < 3; i++) {
ctx.beginPath();
ctx.moveTo(boxes[i].x + boxes[i].w, boxes[i].y + 25);
ctx.lineTo(boxes[i + 1].x, boxes[i + 1].y + 25);
ctx.stroke();
// Arrowhead
ctx.fillStyle = 'rgba(0, 136, 255, 0.6)';
ctx.beginPath();
ctx.moveTo(boxes[i + 1].x, boxes[i + 1].y + 25);
ctx.lineTo(boxes[i + 1].x - 8, boxes[i + 1].y + 20);
ctx.lineTo(boxes[i + 1].x - 8, boxes[i + 1].y + 30);
ctx.fill();
}
// Draw arrows (query pipeline)
ctx.strokeStyle = 'rgba(0, 255, 136, 0.6)';
for (let i = 4; i < 8; i++) {
ctx.beginPath();
ctx.moveTo(boxes[i].x + boxes[i].w, boxes[i].y + 25);
ctx.lineTo(boxes[i + 1].x, boxes[i + 1].y + 25);
ctx.stroke();
ctx.fillStyle = 'rgba(0, 255, 136, 0.6)';
ctx.beginPath();
ctx.moveTo(boxes[i + 1].x, boxes[i + 1].y + 25);
ctx.lineTo(boxes[i + 1].x - 8, boxes[i + 1].y + 20);
ctx.lineTo(boxes[i + 1].x - 8, boxes[i + 1].y + 30);
ctx.fill();
}
// Vector DB connection (vertical)
ctx.strokeStyle = 'rgba(255, 107, 53, 0.6)'; ctx.setLineDash([4, 3]);
ctx.beginPath();
ctx.moveTo(boxes[3].x + 50, boxes[3].y + boxes[3].h);
ctx.lineTo(boxes[6].x + 45, boxes[6].y);
ctx.stroke(); ctx.setLineDash([]);
// Draw boxes
boxes.forEach(b => {
ctx.fillStyle = 'rgba(0, 0, 0, 0.5)';
ctx.fillRect(b.x, b.y, b.w, b.h);
ctx.strokeStyle = b.color; ctx.lineWidth = 2;
ctx.strokeRect(b.x, b.y, b.w, b.h);
ctx.fillStyle = '#e4e6eb'; ctx.font = '11px Arial'; ctx.textAlign = 'center';
ctx.fillText(b.label, b.x + b.w / 2, b.y + 30);
});
// Labels
ctx.font = 'bold 12px Arial'; ctx.textAlign = 'left';
ctx.fillStyle = '#0088ff'; ctx.fillText('Indexing (Offline)', 30, 60);
ctx.fillStyle = '#00ff88'; ctx.fillText('Query (Online)', 30, 170);
}
function drawLoRADiagram(ctx, canvas) {
const w = canvas.width, h = canvas.height;
ctx.fillStyle = '#00c9a7';
ctx.font = 'bold 16px Arial';
ctx.textAlign = 'center';
ctx.fillText('LoRA โ Low-Rank Adaptation', w / 2, 30);
// Original Weight Matrix W (large)
const wX = 50, wY = 60, wW = 120, wH = 120;
ctx.fillStyle = 'rgba(0, 136, 255, 0.15)';
ctx.fillRect(wX, wY, wW, wH);
ctx.strokeStyle = '#0088ff'; ctx.lineWidth = 2;
ctx.strokeRect(wX, wY, wW, wH);
ctx.fillStyle = '#0088ff'; ctx.font = 'bold 14px Arial';
ctx.fillText('W', wX + wW / 2, wY + wH / 2 + 5);
ctx.font = '10px Arial'; ctx.fillStyle = '#b0b7c3';
ctx.fillText('d ร d', wX + wW / 2, wY + wH / 2 + 22);
ctx.fillText('(Frozen)', wX + wW / 2, wY - 8);
// Plus sign
ctx.fillStyle = '#e4e6eb'; ctx.font = 'bold 24px Arial';
ctx.fillText('+', wX + wW + 30, wY + wH / 2 + 8);
// LoRA matrices B and A
const bX = 230, bY = 60, bW = 30, bH = 120;
ctx.fillStyle = 'rgba(0, 255, 136, 0.2)';
ctx.fillRect(bX, bY, bW, bH);
ctx.strokeStyle = '#00ff88'; ctx.lineWidth = 2;
ctx.strokeRect(bX, bY, bW, bH);
ctx.fillStyle = '#00ff88'; ctx.font = 'bold 14px Arial';
ctx.fillText('B', bX + bW / 2, bY + bH / 2 + 5);
ctx.font = '10px Arial'; ctx.fillStyle = '#b0b7c3';
ctx.fillText('dรr', bX + bW / 2, bY - 8);
// ร symbol
ctx.fillStyle = '#e4e6eb'; ctx.font = 'bold 18px Arial';
ctx.fillText('ร', bX + bW + 18, bY + bH / 2 + 5);
const aX = 290, aY = 100, aW = 120, aH = 30;
ctx.fillStyle = 'rgba(255, 107, 53, 0.2)';
ctx.fillRect(aX, aY, aW, aH);
ctx.strokeStyle = '#ff6b35'; ctx.lineWidth = 2;
ctx.strokeRect(aX, aY, aW, aH);
ctx.fillStyle = '#ff6b35'; ctx.font = 'bold 14px Arial';
ctx.fillText('A', aX + aW / 2, aY + aH / 2 + 5);
ctx.font = '10px Arial'; ctx.fillStyle = '#b0b7c3';
ctx.fillText('rรd', aX + aW / 2, aY - 8);
// = sign
ctx.fillStyle = '#e4e6eb'; ctx.font = 'bold 24px Arial';
ctx.fillText('=', aX + aW + 25, wY + wH / 2 + 8);
// Result W'
const rX = 460, rY = 60, rW = 120, rH = 120;
ctx.fillStyle = 'rgba(0, 201, 167, 0.15)';
ctx.fillRect(rX, rY, rW, rH);
ctx.strokeStyle = '#00c9a7'; ctx.lineWidth = 2;
ctx.strokeRect(rX, rY, rW, rH);
ctx.fillStyle = '#00c9a7'; ctx.font = 'bold 14px Arial';
ctx.fillText("W'", rX + rW / 2, rY + rH / 2 + 5);
ctx.font = '10px Arial'; ctx.fillStyle = '#b0b7c3';
ctx.fillText('d ร d', rX + rW / 2, rY + rH / 2 + 22);
ctx.fillText('(Adapted)', rX + rW / 2, rY - 8);
// Stats
ctx.font = '12px Arial'; ctx.textAlign = 'left';
ctx.fillStyle = '#0088ff'; ctx.fillText('Full: 16,777,216 params', 50, 220);
ctx.fillStyle = '#00ff88'; ctx.fillText('LoRA (r=16): 131,072 params', 50, 240);
ctx.fillStyle = '#00c9a7'; ctx.fillText('Savings: 128x reduction (0.78%)', 50, 260);
// Quantization bar chart
ctx.fillStyle = '#e4e6eb'; ctx.font = 'bold 12px Arial'; ctx.textAlign = 'center';
ctx.fillText('Memory: 7B Model', 480, 210);
const bars = [
{ label: 'fp32', val: 28, color: '#ff6b35' },
{ label: 'fp16', val: 14, color: '#ffaa00' },
{ label: 'INT8', val: 7, color: '#0088ff' },
{ label: 'INT4', val: 3.5, color: '#00ff88' }
];
bars.forEach((b, i) => {
const bw = b.val * 4.5;
ctx.fillStyle = b.color;
ctx.fillRect(430, 220 + i * 22, bw, 16);
ctx.fillStyle = '#e4e6eb'; ctx.font = '10px Arial'; ctx.textAlign = 'left';
ctx.fillText(`${b.label}: ${b.val}GB`, 430 + bw + 5, 233 + i * 22);
});
}
initDashboard();
</script>
</body>
</html> |