\uba54\ud0c0-\ub7ec\ub2dd, \ub610\ub294 \uba54\ud0c0 \ud559\uc2b5\uc740 \uae30\uacc4 \ud559\uc2b5\uc758 \ud55c \ubd84\uc57c\ub85c, \ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc774 \ub2e4\ub978 \ud559\uc2b5 \uacfc\uc815\uc744 \ud559\uc2b5\ud558\ub294 \uac83\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774\ub294 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \ub300\ud55c \uc801\uc751\ub825\uc774 \ub6f0\uc5b4\ub09c \ubaa8\ub378\uc744 \ub9cc\ub4dc\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4. \uba54\ud0c0-\ub7ec\ub2dd\uc758 \uc8fc\uc694 \ubaa9\ud45c\ub294 \uacfc\uac70\uc758 \uacbd\ud5d8\uc744 \ubc14\ud0d5\uc73c\ub85c \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc744 \ub354\uc6b1 \ud6a8\uc728\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\ub294 \ubc29\ubc95\uc744 \uc81c\uc2dc\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \uba54\ud0c0-\ub7ec\ub2dd\uc758 \uac1c\ub150\uc744 \uc18c\uac1c\ud558\uace0, \uadf8 \uc911 \ud558\ub098\uc778 MAML(Model-Agnostic Meta-Learning) \uae30\ubc95\uc5d0 \ub300\ud574 \uc0c1\uc138\ud788 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc804\ud1b5\uc801\uc778 \uae30\uacc4 \ud559\uc2b5 \ubaa8\ub378\uc740 \ud2b9\uc815 \uc791\uc5c5\uc5d0 \ub9de\ucd94\uc5b4 \ub9ce\uc740 \uc591\uc758 \ub370\uc774\ud130\ub97c \ud544\uc694\ub85c \ud569\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \ub370\uc774\ud130 \uc218\uc9d1\uc774 \uc5b4\ub835\uac70\ub098 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \uc989\uc2dc \uc801\uc751\ud574\uc57c \ud558\ub294 \uacbd\uc6b0, \uba54\ud0c0-\ub7ec\ub2dd\uc740 \ud2b9\ud788 \uc720\uc6a9\ud569\ub2c8\ub2e4. \uba54\ud0c0-\ub7ec\ub2dd\uc740 \ud559\uc2b5\uc790\uac00 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc744 \ube60\ub974\uac8c \ud559\uc2b5\ud558\uac8c \ub3c4\uc640\uc90d\ub2c8\ub2e4. \uc774\ub294 \ud2b9\ud788 \ub2e4\uc591\ud55c \ud658\uacbd\uc5d0\uc11c \ub2e4\uc591\ud55c \uc791\uc5c5\uc5d0 \uc801\uc751\ud574\uc57c \ud558\ub294 \ub85c\ubd07 \uc81c\uc5b4, \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ub4f1 \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\uba54\ud0c0-\ub7ec\ub2dd \uae30\ubc95\uc740 \uc5ec\ub7ec \uac00\uc9c0\ub85c \ub098\ub258\uba70, \uc8fc\ub85c \ub2e4\uc74c\uacfc \uac19\uc740 \uc138 \uac00\uc9c0 \uce74\ud14c\uace0\ub9ac\ub85c \uad6c\ubd84\ub429\ub2c8\ub2e4:<\/p>\n
MAML\uc740 “\ubaa8\ub378\uc5d0 \uad6c\uc560\ubc1b\uc9c0 \uc54a\ub294 \uba54\ud0c0 \ud559\uc2b5”\uc744 \uc758\ubbf8\ud558\uba70, \uba54\ud0c0-\ub7ec\ub2dd\uc758 \ub300\ud45c\uc801\uc778 \ucd5c\uc801\ud654 \uae30\ubc18 \uae30\ubc95 \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. MAML\uc758 \uc8fc\uc694 \uc544\uc774\ub514\uc5b4\ub294 \ubaa8\ub378\uc774 \uac01\uae30 \ub2e4\ub978 \uc791\uc5c5\uc5d0 \ub300\ud574 \ud6a8\uc728\uc801\uc73c\ub85c \uc801\uc751\ud560 \uc218 \uc788\ub3c4\ub85d \ucd08\uae30\ud654\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. MAML\uc758 \uae30\ubcf8 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
MAML\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \uc218\ud559\uc801 \uc2dd\uc73c\ub85c \ud45c\ud604\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uac00\uc815\ud574 \ubcf4\uc790, \ud2b9\uc815 \uc791\uc5c5 i\uc5d0 \ub300\ud574 \uc190\uc2e4 \ud568\uc218\uac00 \\( \\mathcal{L}_i(\\theta) \\)\ub77c\uba74, MAML\uc740 \ub2e4\uc74c\uc758 \uc218\uc2dd\uc744 \ucd5c\uc801\ud654\ud569\ub2c8\ub2e4:<\/p>\n
\\[
\n \\theta^* = \\arg \\min_\\theta \\sum_{i} \\mathcal{L}_i\\left(\\theta – \\alpha \\nabla \\mathcal{L}_i\\left(\\theta\\right)\\right)
\n \\]<\/p>\n
\uc5ec\uae30\uc11c \\( \\theta \\)\ub294 \ubaa8\ub378\uc758 \ub9e4\uac1c\ubcc0\uc218, \\( \\alpha \\)\ub294 \ud559\uc2b5\ub960, \\( \\nabla \\mathcal{L}_i \\)\ub294 \uc791\uc5c5 i\uc5d0 \ub300\ud55c \uc190\uc2e4 \ud568\uc218\uc758 \uacbd\uc0ac\uc785\ub2c8\ub2e4. MAML\uc740 \ucd08\uae30 \ub9e4\uac1c\ubcc0\uc218 \\( \\theta \\)\ub97c \uc5c5\ub370\uc774\ud2b8\ud558\uc5ec \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \ub300\ud574 \ube60\ub978 \uc801\uc751\uc774 \uac00\ub2a5\ud558\ub3c4\ub85d \ud569\ub2c8\ub2e4.<\/p>\n
MAML\uc758 \uad6c\ud604\uc740 \ubcf4\ud1b5 \ub2e4\uc74c\uc758 \ub2e8\uacc4\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n
\ub2e4\uc74c\uc740 MAML\uc744 \uc774\uc6a9\ud55c \uba54\ud0c0-\ub7ec\ub2dd\uc758 \uac04\ub2e8\ud55c \uad6c\ud604 \uc608\uc2dc\uc785\ub2c8\ub2e4.<\/p>\n
import numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\n\n# \uae30\ubcf8 \ubaa8\ub378 \uc815\uc758\nclass SimpleNN(nn.Module):\n def __init__(self):\n super(SimpleNN, self).__init__()\n self.fc = nn.Sequential(\n nn.Linear(1, 128),\n nn.ReLU(),\n nn.Linear(128, 1)\n )\n\n def forward(self, x):\n return self.fc(x)\n\n# \uc190\uc2e4 \ud568\uc218\uc640 MAML \ub2e8\uacc4 \uad6c\ud604\ndef maml_step(model, data, targets, inner_lr=0.01):\n model.train()\n criterion = nn.MSELoss()\n loss = criterion(model(data), targets)\n \n # \ud30c\ub77c\ubbf8\ud130\uc758 \uacbd\uc0ac \uacc4\uc0b0\n model.zero_grad()\n loss.backward()\n \n # \uacbd\uc0ac \uc218\uc815 \ud6c4 \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n for param in model.parameters():\n param.data -= inner_lr * param.grad.data\n \n return loss.item()\n\n# \uba54\ud0c0 \ud559\uc2b5 \ub8e8\ud504\ndef main():\n model = SimpleNN()\n meta_optimizer = optim.Adam(model.parameters(), lr=0.001)\n \n for meta_step in range(1000):\n meta_train_loss = 0.0\n \n # \uc5ec\ub7ec \uc791\uc5c5\uc5d0 \ub300\ud574 \ubc18\ubcf5\n for task in range(5): # 5\uac1c\uc758 \uc11c\ub85c \ub2e4\ub978 \uc791\uc5c5\n # \uac00\uc0c1 \ub370\uc774\ud130 \uc0dd\uc131\n x = torch.randn(10, 1)\n y = 3 * x + 1 + 0.1 * torch.randn(10, 1) # \uc120\ud615 \uad00\uacc4\n \n # \ube60\ub978 \uc801\uc751\uc744 \uc704\ud55c \ub0b4\ubd80 \ud559\uc2b5\n task_loss = maml_step(model, x, y)\n meta_train_loss += task_loss\n \n # \uba54\ud0c0 \uc190\uc2e4 \uacbd\uacfc \uacc4\uc0b0 \ubc0f \ucd5c\uc801\ud654\n meta_optimizer.zero_grad()\n (meta_train_loss \/ 5).backward()\n meta_optimizer.step()\n \n if meta_step % 100 == 0:\n print(f'Meta Step: {meta_step}, Meta Train Loss: {meta_train_loss \/ 5}')\n\nif __name__ == \"__main__\":\n main()\n<\/code><\/pre>\nMAML\uc758 \uc7a5\uc810 \ubc0f \ub2e8\uc810<\/h3>\n
MAML\uc758 \uc8fc\uc694 \uc7a5\uc810\uc740 \ubaa8\ub378\uc774 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \ub300\ud574 \ube60\ub974\uac8c \uc801\uc751\ud560 \uc218 \uc788\ub3c4\ub85d \ucd08\uae30\ud654\ud560 \uc218 \uc788\ub2e4\ub294 \uc810\uc785\ub2c8\ub2e4. \ucd08\uae30 \ub9e4\uac1c\ubcc0\uc218\uac00 \ub2e4\uc591\ud55c \uc791\uc5c5\uc758 \uc5ec\ub7ec \uce21\uba74\uc744 \ud3ec\ucc29\ud560 \uc218 \uc788\uc73c\uba70, \uc801\uc740 \ub370\uc774\ud130\ub85c\ub3c4 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uadf8\ub7ec\ub098 MAML\uc740 \uba87 \uac00\uc9c0 \ub2e8\uc810 \ub610\ud55c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uccab\uc9f8, \ubcf5\uc7a1\ud55c \ubaa8\ub378 \uad6c\uc870\uc5d0\uc11c\ub294 \ud559\uc2b5 \uacfc\uc815\uc774 \ub290\ub824\uc9c8 \uc218 \uc788\uc73c\uba70, \ub458\uc9f8, \uc801\uc808\ud55c \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130 \uc120\ud0dd\uc774 \uacb0\uacfc\uc5d0 \ud06c\uac8c \uc601\ud5a5\uc744 \ubbf8\uce60 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc751\uc6a9 \ubd84\uc57c<\/h2>\n
MAML\uc744 \ud3ec\ud568\ud55c \uba54\ud0c0-\ub7ec\ub2dd \uae30\ubc95\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\ubc1c\ud788 \uc5f0\uad6c\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8 \uc911 \uba87 \uac00\uc9c0 \ub300\ud45c\uc801\uc778 \uc751\uc6a9 \ubd84\uc57c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n